[ { "id": "agibot_683_868633-train-bf_001397-af_001824-plus-0000341", "task": "Serve the meal", "image": [ "images/agibotworld/683/868698/cam_high/frame_000000.jpg", "images/agibotworld/683/868698/cam_high/frame_001879.jpg", "images/agibotworld/683/868633/cam_high/frame_001397.jpg", "images/agibotworld/683/868633/cam_left_wrist/frame_001397.jpg", "images/agibotworld/683/868633/cam_right_wrist/frame_001397.jpg", "images/agibotworld/683/868633/cam_high/frame_001824.jpg", "images/agibotworld/683/868633/cam_left_wrist/frame_001824.jpg", "images/agibotworld/683/868633/cam_right_wrist/frame_001824.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Serve the meal`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Serve the meal`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 42.9, "hop_absolute_value": 42.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_375_694775-train-bf_000737-af_000822-plus-0000723", "task": "Brew tea", "image": [ "images/agibotworld/375/695466/cam_high/frame_000000.jpg", "images/agibotworld/375/695466/cam_high/frame_001814.jpg", "images/agibotworld/375/694775/cam_high/frame_000737.jpg", "images/agibotworld/375/694775/cam_left_wrist/frame_000737.jpg", "images/agibotworld/375/694775/cam_right_wrist/frame_000737.jpg", "images/agibotworld/375/694775/cam_high/frame_000822.jpg", "images/agibotworld/375/694775/cam_left_wrist/frame_000822.jpg", "images/agibotworld/375/694775/cam_right_wrist/frame_000822.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Brew tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Brew tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 6.7, "hop_absolute_value": 6.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_431_782894-train-bf_000199-af_001933-plus-0000560", "task": "Brew coffee with a capsule machine", "image": [ "images/agibotworld/431/787084/cam_high/frame_000000.jpg", "images/agibotworld/431/787084/cam_high/frame_001467.jpg", "images/agibotworld/431/782894/cam_high/frame_000199.jpg", "images/agibotworld/431/782894/cam_left_wrist/frame_000199.jpg", "images/agibotworld/431/782894/cam_right_wrist/frame_000199.jpg", "images/agibotworld/431/782894/cam_high/frame_001933.jpg", "images/agibotworld/431/782894/cam_left_wrist/frame_001933.jpg", "images/agibotworld/431/782894/cam_right_wrist/frame_001933.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Brew coffee with a capsule machine`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Brew coffee with a capsule machine`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_388_690677-train-bf_000076-af_000384-plus-0000785", "task": "Pickup in the supermarket", "image": [ "images/agibotworld/388/691251/cam_high/frame_000039.jpg", "images/agibotworld/388/691251/cam_high/frame_001435.jpg", "images/agibotworld/388/690677/cam_high/frame_000076.jpg", "images/agibotworld/388/690677/cam_left_wrist/frame_000076.jpg", "images/agibotworld/388/690677/cam_right_wrist/frame_000076.jpg", "images/agibotworld/388/690677/cam_high/frame_000384.jpg", "images/agibotworld/388/690677/cam_left_wrist/frame_000384.jpg", "images/agibotworld/388/690677/cam_right_wrist/frame_000384.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pickup in the supermarket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pickup in the supermarket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 22.9, "hop_absolute_value": 22.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_609_832750-train-bf_000093-af_000140-plus-0000470", "task": "Restock the hanging basket area", "image": [ "images/agibotworld/609/832750/cam_high/frame_000000.jpg", "images/agibotworld/609/832750/cam_high/frame_001012.jpg", "images/agibotworld/609/832750/cam_high/frame_000093.jpg", "images/agibotworld/609/832750/cam_left_wrist/frame_000093.jpg", "images/agibotworld/609/832750/cam_right_wrist/frame_000093.jpg", "images/agibotworld/609/832750/cam_high/frame_000140.jpg", "images/agibotworld/609/832750/cam_left_wrist/frame_000140.jpg", "images/agibotworld/609/832750/cam_right_wrist/frame_000140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Restock the hanging basket area`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Restock the hanging basket area`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.5, "hop_absolute_value": 4.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_580_862866-train-bf_000592-af_004163-plus-0000447", "task": "Make juice", "image": [ "images/agibotworld/580/867528/cam_high/frame_000000.jpg", "images/agibotworld/580/867528/cam_high/frame_007072.jpg", "images/agibotworld/580/862866/cam_high/frame_000592.jpg", "images/agibotworld/580/862866/cam_left_wrist/frame_000592.jpg", "images/agibotworld/580/862866/cam_right_wrist/frame_000592.jpg", "images/agibotworld/580/862866/cam_high/frame_004163.jpg", "images/agibotworld/580/862866/cam_left_wrist/frame_004163.jpg", "images/agibotworld/580/862866/cam_right_wrist/frame_004163.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make juice`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make juice`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 56.2, "hop_absolute_value": 56.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_621_880091-train-bf_002112-af_003938-plus-0000497", "task": "Add the seasoning to the pot.", "image": [ "images/agibotworld/621/872985/cam_high/frame_000126.jpg", "images/agibotworld/621/872985/cam_high/frame_005272.jpg", "images/agibotworld/621/880091/cam_high/frame_002112.jpg", "images/agibotworld/621/880091/cam_left_wrist/frame_002112.jpg", "images/agibotworld/621/880091/cam_right_wrist/frame_002112.jpg", "images/agibotworld/621/880091/cam_high/frame_003938.jpg", "images/agibotworld/621/880091/cam_left_wrist/frame_003938.jpg", "images/agibotworld/621/880091/cam_right_wrist/frame_003938.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Add the seasoning to the pot.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Add the seasoning to the pot.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 73.9, "hop_absolute_value": 73.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_774_925510-train-bf_000107-af_000780-plus-0000294", "task": "Insert the key and open the door", "image": [ "images/agibotworld/774/889716/cam_high/frame_000000.jpg", "images/agibotworld/774/889716/cam_high/frame_000785.jpg", "images/agibotworld/774/925510/cam_high/frame_000107.jpg", "images/agibotworld/774/925510/cam_left_wrist/frame_000107.jpg", "images/agibotworld/774/925510/cam_right_wrist/frame_000107.jpg", "images/agibotworld/774/925510/cam_high/frame_000780.jpg", "images/agibotworld/774/925510/cam_left_wrist/frame_000780.jpg", "images/agibotworld/774/925510/cam_right_wrist/frame_000780.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert the key and open the door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert the key and open the door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 43.5, "hop_absolute_value": 43.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_487_720041-train-bf_002090-af_003768-plus-0000545", "task": "Remove bottled water from the carton and arrange neatly on the table.", "image": [ "images/agibotworld/487/719828/cam_high/frame_000039.jpg", "images/agibotworld/487/719828/cam_high/frame_003576.jpg", "images/agibotworld/487/720041/cam_high/frame_002090.jpg", "images/agibotworld/487/720041/cam_left_wrist/frame_002090.jpg", "images/agibotworld/487/720041/cam_right_wrist/frame_002090.jpg", "images/agibotworld/487/720041/cam_high/frame_003768.jpg", "images/agibotworld/487/720041/cam_left_wrist/frame_003768.jpg", "images/agibotworld/487/720041/cam_right_wrist/frame_003768.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove bottled water from the carton and arrange neatly on the table.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove bottled water from the carton and arrange neatly on the table.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 91.9, "hop_absolute_value": 91.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_521_763453-train-bf_000943-af_001206-plus-0000464", "task": "Dispose of the trash on the desk.", "image": [ "images/agibotworld/521/757938/cam_high/frame_000033.jpg", "images/agibotworld/521/757938/cam_high/frame_001643.jpg", "images/agibotworld/521/763453/cam_high/frame_000943.jpg", "images/agibotworld/521/763453/cam_left_wrist/frame_000943.jpg", "images/agibotworld/521/763453/cam_right_wrist/frame_000943.jpg", "images/agibotworld/521/763453/cam_high/frame_001206.jpg", "images/agibotworld/521/763453/cam_left_wrist/frame_001206.jpg", "images/agibotworld/521/763453/cam_right_wrist/frame_001206.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of the trash on the desk.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of the trash on the desk.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_480_728736-train-bf_000974-af_001066-plus-0000949", "task": "Make sandwiches with salad dressing 1213 edition", "image": [ "images/agibotworld/480/728736/cam_high/frame_000000.jpg", "images/agibotworld/480/728736/cam_high/frame_002388.jpg", "images/agibotworld/480/728736/cam_high/frame_000974.jpg", "images/agibotworld/480/728736/cam_left_wrist/frame_000974.jpg", "images/agibotworld/480/728736/cam_right_wrist/frame_000974.jpg", "images/agibotworld/480/728736/cam_high/frame_001066.jpg", "images/agibotworld/480/728736/cam_left_wrist/frame_001066.jpg", "images/agibotworld/480/728736/cam_right_wrist/frame_001066.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make sandwiches with salad dressing 1213 edition`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make sandwiches with salad dressing 1213 edition`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 7.7, "hop_absolute_value": 7.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_782_914994-train-bf_000604-af_000690-plus-0000784", "task": "Tidy the bar counter", "image": [ "images/agibotworld/782/914693/cam_high/frame_000064.jpg", "images/agibotworld/782/914693/cam_high/frame_001915.jpg", "images/agibotworld/782/914994/cam_high/frame_000604.jpg", "images/agibotworld/782/914994/cam_left_wrist/frame_000604.jpg", "images/agibotworld/782/914994/cam_right_wrist/frame_000604.jpg", "images/agibotworld/782/914994/cam_high/frame_000690.jpg", "images/agibotworld/782/914994/cam_left_wrist/frame_000690.jpg", "images/agibotworld/782/914994/cam_right_wrist/frame_000690.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Tidy the bar counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Tidy the bar counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.4, "hop_absolute_value": 5.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_550_821460-train-bf_000265-af_000362-plus-0000654", "task": "Peel fruits and vegetables", "image": [ "images/agibotworld/550/832819/cam_high/frame_000046.jpg", "images/agibotworld/550/832819/cam_high/frame_001720.jpg", "images/agibotworld/550/821460/cam_high/frame_000265.jpg", "images/agibotworld/550/821460/cam_left_wrist/frame_000265.jpg", "images/agibotworld/550/821460/cam_right_wrist/frame_000265.jpg", "images/agibotworld/550/821460/cam_high/frame_000362.jpg", "images/agibotworld/550/821460/cam_left_wrist/frame_000362.jpg", "images/agibotworld/550/821460/cam_right_wrist/frame_000362.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Peel fruits and vegetables`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Peel fruits and vegetables`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.7, "hop_absolute_value": 10.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_454_692968-train-bf_002827-af_003299-plus-0000565", "task": "Make a sandwich", "image": [ "images/agibotworld/454/688896/cam_high/frame_000035.jpg", "images/agibotworld/454/688896/cam_high/frame_003001.jpg", "images/agibotworld/454/692968/cam_high/frame_002827.jpg", "images/agibotworld/454/692968/cam_left_wrist/frame_002827.jpg", "images/agibotworld/454/692968/cam_right_wrist/frame_002827.jpg", "images/agibotworld/454/692968/cam_high/frame_003299.jpg", "images/agibotworld/454/692968/cam_left_wrist/frame_003299.jpg", "images/agibotworld/454/692968/cam_right_wrist/frame_003299.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make a sandwich`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make a sandwich`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 93.8, "hop_absolute_value": 93.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_543_808768-train-bf_000273-af_000304-plus-0000476", "task": "Packing washing detergent", "image": [ "images/agibotworld/543/808768/cam_high/frame_000116.jpg", "images/agibotworld/543/808768/cam_high/frame_000761.jpg", "images/agibotworld/543/808768/cam_high/frame_000273.jpg", "images/agibotworld/543/808768/cam_left_wrist/frame_000273.jpg", "images/agibotworld/543/808768/cam_right_wrist/frame_000273.jpg", "images/agibotworld/543/808768/cam_high/frame_000304.jpg", "images/agibotworld/543/808768/cam_left_wrist/frame_000304.jpg", "images/agibotworld/543/808768/cam_right_wrist/frame_000304.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Packing washing detergent`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Packing washing detergent`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.3, "hop_absolute_value": 5.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_452_715411-train-bf_000333-af_000376-plus-0000277", "task": "Slice a lemon to make lemon water.", "image": [ "images/agibotworld/452/714273/cam_high/frame_000020.jpg", "images/agibotworld/452/714273/cam_high/frame_001388.jpg", "images/agibotworld/452/715411/cam_high/frame_000333.jpg", "images/agibotworld/452/715411/cam_left_wrist/frame_000333.jpg", "images/agibotworld/452/715411/cam_right_wrist/frame_000333.jpg", "images/agibotworld/452/715411/cam_high/frame_000376.jpg", "images/agibotworld/452/715411/cam_left_wrist/frame_000376.jpg", "images/agibotworld/452/715411/cam_right_wrist/frame_000376.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Slice a lemon to make lemon water.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Slice a lemon to make lemon water.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.9, "hop_absolute_value": 5.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_708_900178-train-bf_000975-af_005027-plus-0000566", "task": "Place name tags", "image": [ "images/agibotworld/708/910844/cam_high/frame_000000.jpg", "images/agibotworld/708/910844/cam_high/frame_004089.jpg", "images/agibotworld/708/900178/cam_high/frame_000975.jpg", "images/agibotworld/708/900178/cam_left_wrist/frame_000975.jpg", "images/agibotworld/708/900178/cam_right_wrist/frame_000975.jpg", "images/agibotworld/708/900178/cam_high/frame_005027.jpg", "images/agibotworld/708/900178/cam_left_wrist/frame_005027.jpg", "images/agibotworld/708/900178/cam_right_wrist/frame_005027.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place name tags`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place name tags`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 89.7, "hop_absolute_value": 89.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_372_664533-train-bf_000731-af_000925-plus-0000307", "task": "Packing in the supermarket", "image": [ "images/agibotworld/372/664533/cam_high/frame_000019.jpg", "images/agibotworld/372/664533/cam_high/frame_002140.jpg", "images/agibotworld/372/664533/cam_high/frame_000731.jpg", "images/agibotworld/372/664533/cam_left_wrist/frame_000731.jpg", "images/agibotworld/372/664533/cam_right_wrist/frame_000731.jpg", "images/agibotworld/372/664533/cam_high/frame_000925.jpg", "images/agibotworld/372/664533/cam_left_wrist/frame_000925.jpg", "images/agibotworld/372/664533/cam_right_wrist/frame_000925.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Packing in the supermarket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Packing in the supermarket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_504_769603-train-bf_000849-af_001043-plus-0000351", "task": "Restock supermarket snacks", "image": [ "images/agibotworld/504/727913/cam_high/frame_000000.jpg", "images/agibotworld/504/727913/cam_high/frame_002251.jpg", "images/agibotworld/504/769603/cam_high/frame_000849.jpg", "images/agibotworld/504/769603/cam_left_wrist/frame_000849.jpg", "images/agibotworld/504/769603/cam_right_wrist/frame_000849.jpg", "images/agibotworld/504/769603/cam_high/frame_001043.jpg", "images/agibotworld/504/769603/cam_left_wrist/frame_001043.jpg", "images/agibotworld/504/769603/cam_right_wrist/frame_001043.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Restock supermarket snacks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Restock supermarket snacks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 42.9, "hop_absolute_value": 42.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_582_803488-train-bf_000381-af_000591-plus-0000317", "task": "Discard the trash on the coffee table.", "image": [ "images/agibotworld/582/803503/cam_high/frame_000000.jpg", "images/agibotworld/582/803503/cam_high/frame_000664.jpg", "images/agibotworld/582/803488/cam_high/frame_000381.jpg", "images/agibotworld/582/803488/cam_left_wrist/frame_000381.jpg", "images/agibotworld/582/803488/cam_right_wrist/frame_000381.jpg", "images/agibotworld/582/803488/cam_high/frame_000591.jpg", "images/agibotworld/582/803488/cam_left_wrist/frame_000591.jpg", "images/agibotworld/582/803488/cam_right_wrist/frame_000591.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Discard the trash on the coffee table.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Discard the trash on the coffee table.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 58.3, "hop_absolute_value": 58.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_547_843250-train-bf_002496-af_002574-plus-0000914", "task": "Sort light and dark clothes", "image": [ "images/agibotworld/547/843250/cam_high/frame_000000.jpg", "images/agibotworld/547/843250/cam_high/frame_003381.jpg", "images/agibotworld/547/843250/cam_high/frame_002496.jpg", "images/agibotworld/547/843250/cam_left_wrist/frame_002496.jpg", "images/agibotworld/547/843250/cam_right_wrist/frame_002496.jpg", "images/agibotworld/547/843250/cam_high/frame_002574.jpg", "images/agibotworld/547/843250/cam_left_wrist/frame_002574.jpg", "images/agibotworld/547/843250/cam_right_wrist/frame_002574.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort light and dark clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort light and dark clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 6.2, "hop_absolute_value": 6.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_542_775271-train-bf_000491-af_000538-plus-0000589", "task": "Boil water with a kettle", "image": [ "images/agibotworld/542/775271/cam_high/frame_000000.jpg", "images/agibotworld/542/775271/cam_high/frame_001632.jpg", "images/agibotworld/542/775271/cam_high/frame_000491.jpg", "images/agibotworld/542/775271/cam_left_wrist/frame_000491.jpg", "images/agibotworld/542/775271/cam_right_wrist/frame_000491.jpg", "images/agibotworld/542/775271/cam_high/frame_000538.jpg", "images/agibotworld/542/775271/cam_left_wrist/frame_000538.jpg", "images/agibotworld/542/775271/cam_right_wrist/frame_000538.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Boil water with a kettle`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Boil water with a kettle`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.6, "hop_absolute_value": 3.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_694_881865-train-bf_000430-af_001945-plus-0000547", "task": "Twist the bottle cap", "image": [ "images/agibotworld/694/882504/cam_high/frame_000000.jpg", "images/agibotworld/694/882504/cam_high/frame_002115.jpg", "images/agibotworld/694/881865/cam_high/frame_000430.jpg", "images/agibotworld/694/881865/cam_left_wrist/frame_000430.jpg", "images/agibotworld/694/881865/cam_right_wrist/frame_000430.jpg", "images/agibotworld/694/881865/cam_high/frame_001945.jpg", "images/agibotworld/694/881865/cam_left_wrist/frame_001945.jpg", "images/agibotworld/694/881865/cam_right_wrist/frame_001945.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 93.1, "hop_absolute_value": 93.1, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_543_823616-train-bf_000741-af_000972-plus-0000428", "task": "Packing washing detergent", "image": [ "images/agibotworld/543/823616/cam_high/frame_000078.jpg", "images/agibotworld/543/823616/cam_high/frame_000972.jpg", "images/agibotworld/543/823616/cam_high/frame_000741.jpg", "images/agibotworld/543/823616/cam_left_wrist/frame_000741.jpg", "images/agibotworld/543/823616/cam_right_wrist/frame_000741.jpg", "images/agibotworld/543/823616/cam_high/frame_000972.jpg", "images/agibotworld/543/823616/cam_left_wrist/frame_000972.jpg", "images/agibotworld/543/823616/cam_right_wrist/frame_000972.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Packing washing detergent`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Packing washing detergent`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_424_690833-train-bf_001188-af_001557-plus-0000525", "task": "Clear the countertop waste", "image": [ "images/agibotworld/424/700548/cam_high/frame_000000.jpg", "images/agibotworld/424/700548/cam_high/frame_001767.jpg", "images/agibotworld/424/690833/cam_high/frame_001188.jpg", "images/agibotworld/424/690833/cam_left_wrist/frame_001188.jpg", "images/agibotworld/424/690833/cam_right_wrist/frame_001188.jpg", "images/agibotworld/424/690833/cam_high/frame_001557.jpg", "images/agibotworld/424/690833/cam_left_wrist/frame_001557.jpg", "images/agibotworld/424/690833/cam_right_wrist/frame_001557.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Clear the countertop waste`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Clear the countertop waste`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 88.9, "hop_absolute_value": 88.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_465_738953-train-bf_000920-af_002452-plus-0000546", "task": "Wash clothes in the washing machine.", "image": [ "images/agibotworld/465/742238/cam_high/frame_000000.jpg", "images/agibotworld/465/742238/cam_high/frame_002820.jpg", "images/agibotworld/465/738953/cam_high/frame_000920.jpg", "images/agibotworld/465/738953/cam_left_wrist/frame_000920.jpg", "images/agibotworld/465/738953/cam_right_wrist/frame_000920.jpg", "images/agibotworld/465/738953/cam_high/frame_002452.jpg", "images/agibotworld/465/738953/cam_left_wrist/frame_002452.jpg", "images/agibotworld/465/738953/cam_right_wrist/frame_002452.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wash clothes in the washing machine.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wash clothes in the washing machine.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 89.7, "hop_absolute_value": 89.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_735_903144-train-bf_001082-af_001119-plus-0000750", "task": "Roll away stains with a lint roller", "image": [ "images/agibotworld/735/908091/cam_high/frame_000000.jpg", "images/agibotworld/735/908091/cam_high/frame_001798.jpg", "images/agibotworld/735/903144/cam_high/frame_001082.jpg", "images/agibotworld/735/903144/cam_left_wrist/frame_001082.jpg", "images/agibotworld/735/903144/cam_right_wrist/frame_001082.jpg", "images/agibotworld/735/903144/cam_high/frame_001119.jpg", "images/agibotworld/735/903144/cam_left_wrist/frame_001119.jpg", "images/agibotworld/735/903144/cam_right_wrist/frame_001119.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Roll away stains with a lint roller`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Roll away stains with a lint roller`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 6.7, "hop_absolute_value": 6.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_366_657006-train-bf_000134-af_001115-plus-0000498", "task": "Sort food", "image": [ "images/agibotworld/366/669950/cam_high/frame_000025.jpg", "images/agibotworld/366/669950/cam_high/frame_001145.jpg", "images/agibotworld/366/657006/cam_high/frame_000134.jpg", "images/agibotworld/366/657006/cam_left_wrist/frame_000134.jpg", "images/agibotworld/366/657006/cam_right_wrist/frame_000134.jpg", "images/agibotworld/366/657006/cam_high/frame_001115.jpg", "images/agibotworld/366/657006/cam_left_wrist/frame_001115.jpg", "images/agibotworld/366/657006/cam_right_wrist/frame_001115.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort food`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort food`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 88.9, "hop_absolute_value": 88.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_359_693937-train-bf_000048-af_000559-plus-0000381", "task": "Sort in the warehouse", "image": [ "images/agibotworld/359/693937/cam_high/frame_000000.jpg", "images/agibotworld/359/693937/cam_high/frame_000833.jpg", "images/agibotworld/359/693937/cam_high/frame_000048.jpg", "images/agibotworld/359/693937/cam_left_wrist/frame_000048.jpg", "images/agibotworld/359/693937/cam_right_wrist/frame_000048.jpg", "images/agibotworld/359/693937/cam_high/frame_000559.jpg", "images/agibotworld/359/693937/cam_left_wrist/frame_000559.jpg", "images/agibotworld/359/693937/cam_right_wrist/frame_000559.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort in the warehouse`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort in the warehouse`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 58.6, "hop_absolute_value": 58.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_525_766236-train-bf_000868-af_001099-plus-0000911", "task": "Place cutlery in the restaurant", "image": [ "images/agibotworld/525/771244/cam_high/frame_000000.jpg", "images/agibotworld/525/771244/cam_high/frame_001628.jpg", "images/agibotworld/525/766236/cam_high/frame_000868.jpg", "images/agibotworld/525/766236/cam_left_wrist/frame_000868.jpg", "images/agibotworld/525/766236/cam_right_wrist/frame_000868.jpg", "images/agibotworld/525/766236/cam_high/frame_001099.jpg", "images/agibotworld/525/766236/cam_left_wrist/frame_001099.jpg", "images/agibotworld/525/766236/cam_right_wrist/frame_001099.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place cutlery in the restaurant`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place cutlery in the restaurant`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.1, "hop_absolute_value": 12.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_366_670024-train-bf_000555-af_000902-plus-0000363", "task": "Sort food", "image": [ "images/agibotworld/366/669950/cam_high/frame_000025.jpg", "images/agibotworld/366/669950/cam_high/frame_001145.jpg", "images/agibotworld/366/670024/cam_high/frame_000555.jpg", "images/agibotworld/366/670024/cam_left_wrist/frame_000555.jpg", "images/agibotworld/366/670024/cam_right_wrist/frame_000555.jpg", "images/agibotworld/366/670024/cam_high/frame_000902.jpg", "images/agibotworld/366/670024/cam_left_wrist/frame_000902.jpg", "images/agibotworld/366/670024/cam_right_wrist/frame_000902.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort food`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort food`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 53.3, "hop_absolute_value": 53.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_356_654636-train-bf_000415-af_000706-plus-0000453", "task": "Packing in the supermarket", "image": [ "images/agibotworld/356/649079/cam_high/frame_000008.jpg", "images/agibotworld/356/649079/cam_high/frame_000854.jpg", "images/agibotworld/356/654636/cam_high/frame_000415.jpg", "images/agibotworld/356/654636/cam_left_wrist/frame_000415.jpg", "images/agibotworld/356/654636/cam_right_wrist/frame_000415.jpg", "images/agibotworld/356/654636/cam_high/frame_000706.jpg", "images/agibotworld/356/654636/cam_left_wrist/frame_000706.jpg", "images/agibotworld/356/654636/cam_right_wrist/frame_000706.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Packing in the supermarket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Packing in the supermarket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 82.4, "hop_absolute_value": 82.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_440_747804-train-bf_000202-af_000486-plus-0000891", "task": "Iron clothes", "image": [ "images/agibotworld/440/696613/cam_high/frame_000000.jpg", "images/agibotworld/440/696613/cam_high/frame_002346.jpg", "images/agibotworld/440/747804/cam_high/frame_000202.jpg", "images/agibotworld/440/747804/cam_left_wrist/frame_000202.jpg", "images/agibotworld/440/747804/cam_right_wrist/frame_000202.jpg", "images/agibotworld/440/747804/cam_high/frame_000486.jpg", "images/agibotworld/440/747804/cam_left_wrist/frame_000486.jpg", "images/agibotworld/440/747804/cam_right_wrist/frame_000486.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Iron clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Iron clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 13.3, "hop_absolute_value": 13.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_508_763112-train-bf_001444-af_001473-plus-0000698", "task": "Place items in the bag.", "image": [ "images/agibotworld/508/763112/cam_high/frame_000000.jpg", "images/agibotworld/508/763112/cam_high/frame_002031.jpg", "images/agibotworld/508/763112/cam_high/frame_001444.jpg", "images/agibotworld/508/763112/cam_left_wrist/frame_001444.jpg", "images/agibotworld/508/763112/cam_right_wrist/frame_001444.jpg", "images/agibotworld/508/763112/cam_high/frame_001473.jpg", "images/agibotworld/508/763112/cam_left_wrist/frame_001473.jpg", "images/agibotworld/508/763112/cam_right_wrist/frame_001473.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place items in the bag.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place items in the bag.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.9, "hop_absolute_value": 5.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_352_662346-train-bf_000670-af_001276-plus-0000441", "task": "Open the fridge to get food.", "image": [ "images/agibotworld/352/662346/cam_high/frame_000012.jpg", "images/agibotworld/352/662346/cam_high/frame_001315.jpg", "images/agibotworld/352/662346/cam_high/frame_000670.jpg", "images/agibotworld/352/662346/cam_left_wrist/frame_000670.jpg", "images/agibotworld/352/662346/cam_right_wrist/frame_000670.jpg", "images/agibotworld/352/662346/cam_high/frame_001276.jpg", "images/agibotworld/352/662346/cam_left_wrist/frame_001276.jpg", "images/agibotworld/352/662346/cam_right_wrist/frame_001276.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the fridge to get food.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the fridge to get food.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 93.8, "hop_absolute_value": 93.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_510_767554-train-bf_000071-af_001114-plus-0000383", "task": "Stack dishcloth on the kitchen countertop", "image": [ "images/agibotworld/510/746437/cam_high/frame_000000.jpg", "images/agibotworld/510/746437/cam_high/frame_002768.jpg", "images/agibotworld/510/767554/cam_high/frame_000071.jpg", "images/agibotworld/510/767554/cam_left_wrist/frame_000071.jpg", "images/agibotworld/510/767554/cam_right_wrist/frame_000071.jpg", "images/agibotworld/510/767554/cam_high/frame_001114.jpg", "images/agibotworld/510/767554/cam_left_wrist/frame_001114.jpg", "images/agibotworld/510/767554/cam_right_wrist/frame_001114.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Stack dishcloth on the kitchen countertop`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Stack dishcloth on the kitchen countertop`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 55.2, "hop_absolute_value": 55.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_547_842961-train-bf_001592-af_002567-plus-0000433", "task": "Sort light and dark clothes", "image": [ "images/agibotworld/547/845046/cam_high/frame_000000.jpg", "images/agibotworld/547/845046/cam_high/frame_003406.jpg", "images/agibotworld/547/842961/cam_high/frame_001592.jpg", "images/agibotworld/547/842961/cam_left_wrist/frame_001592.jpg", "images/agibotworld/547/842961/cam_right_wrist/frame_001592.jpg", "images/agibotworld/547/842961/cam_high/frame_002567.jpg", "images/agibotworld/547/842961/cam_left_wrist/frame_002567.jpg", "images/agibotworld/547/842961/cam_right_wrist/frame_002567.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort light and dark clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort light and dark clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 63.0, "hop_absolute_value": 63.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_741_879705-train-bf_000746-af_000806-plus-0000637", "task": "Pack the fruits", "image": [ "images/agibotworld/741/879705/cam_high/frame_000000.jpg", "images/agibotworld/741/879705/cam_high/frame_001401.jpg", "images/agibotworld/741/879705/cam_high/frame_000746.jpg", "images/agibotworld/741/879705/cam_left_wrist/frame_000746.jpg", "images/agibotworld/741/879705/cam_right_wrist/frame_000746.jpg", "images/agibotworld/741/879705/cam_high/frame_000806.jpg", "images/agibotworld/741/879705/cam_left_wrist/frame_000806.jpg", "images/agibotworld/741/879705/cam_right_wrist/frame_000806.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pack the fruits`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pack the fruits`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.5, "hop_absolute_value": 10.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_716_868536-train-bf_000375-af_000545-plus-0000925", "task": "Stamp the document and place the reimbursement form in the reimbursement box.", "image": [ "images/agibotworld/716/915968/cam_high/frame_000000.jpg", "images/agibotworld/716/915968/cam_high/frame_001334.jpg", "images/agibotworld/716/868536/cam_high/frame_000375.jpg", "images/agibotworld/716/868536/cam_left_wrist/frame_000375.jpg", "images/agibotworld/716/868536/cam_right_wrist/frame_000375.jpg", "images/agibotworld/716/868536/cam_high/frame_000545.jpg", "images/agibotworld/716/868536/cam_left_wrist/frame_000545.jpg", "images/agibotworld/716/868536/cam_right_wrist/frame_000545.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Stamp the document and place the reimbursement form in the reimbursement box.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Stamp the document and place the reimbursement form in the reimbursement box.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_547_844931-train-bf_002375-af_003309-plus-0000563", "task": "Sort light and dark clothes", "image": [ "images/agibotworld/547/842295/cam_high/frame_000000.jpg", "images/agibotworld/547/842295/cam_high/frame_003369.jpg", "images/agibotworld/547/844931/cam_high/frame_002375.jpg", "images/agibotworld/547/844931/cam_left_wrist/frame_002375.jpg", "images/agibotworld/547/844931/cam_right_wrist/frame_002375.jpg", "images/agibotworld/547/844931/cam_high/frame_003309.jpg", "images/agibotworld/547/844931/cam_left_wrist/frame_003309.jpg", "images/agibotworld/547/844931/cam_right_wrist/frame_003309.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort light and dark clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort light and dark clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 93.8, "hop_absolute_value": 93.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_327_648756-train-bf_000077-af_000888-plus-0000427", "task": "Pickup items in the supermarket", "image": [ "images/agibotworld/327/649845/cam_high/frame_000000.jpg", "images/agibotworld/327/649845/cam_high/frame_001283.jpg", "images/agibotworld/327/648756/cam_high/frame_000077.jpg", "images/agibotworld/327/648756/cam_left_wrist/frame_000077.jpg", "images/agibotworld/327/648756/cam_right_wrist/frame_000077.jpg", "images/agibotworld/327/648756/cam_high/frame_000888.jpg", "images/agibotworld/327/648756/cam_left_wrist/frame_000888.jpg", "images/agibotworld/327/648756/cam_right_wrist/frame_000888.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pickup items in the supermarket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pickup items in the supermarket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 61.8, "hop_absolute_value": 61.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_563_839342-train-bf_000253-af_000480-plus-0000677", "task": "Remove the baked dessert from the oven.", "image": [ "images/agibotworld/563/863704/cam_high/frame_000000.jpg", "images/agibotworld/563/863704/cam_high/frame_002630.jpg", "images/agibotworld/563/839342/cam_high/frame_000253.jpg", "images/agibotworld/563/839342/cam_left_wrist/frame_000253.jpg", "images/agibotworld/563/839342/cam_right_wrist/frame_000253.jpg", "images/agibotworld/563/839342/cam_high/frame_000480.jpg", "images/agibotworld/563/839342/cam_left_wrist/frame_000480.jpg", "images/agibotworld/563/839342/cam_right_wrist/frame_000480.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the baked dessert from the oven.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the baked dessert from the oven.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.9, "hop_absolute_value": 2.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_452_719134-train-bf_000595-af_000794-plus-0000721", "task": "Slice a lemon to make lemon water.", "image": [ "images/agibotworld/452/731052/cam_high/frame_000000.jpg", "images/agibotworld/452/731052/cam_high/frame_001370.jpg", "images/agibotworld/452/719134/cam_high/frame_000595.jpg", "images/agibotworld/452/719134/cam_left_wrist/frame_000595.jpg", "images/agibotworld/452/719134/cam_right_wrist/frame_000595.jpg", "images/agibotworld/452/719134/cam_high/frame_000794.jpg", "images/agibotworld/452/719134/cam_left_wrist/frame_000794.jpg", "images/agibotworld/452/719134/cam_right_wrist/frame_000794.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Slice a lemon to make lemon water.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Slice a lemon to make lemon water.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 28.6, "hop_absolute_value": 28.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_751_908652-train-bf_000239-af_000928-plus-0000494", "task": "Slice the noodles", "image": [ "images/agibotworld/751/902247/cam_high/frame_000000.jpg", "images/agibotworld/751/902247/cam_high/frame_001578.jpg", "images/agibotworld/751/908652/cam_high/frame_000239.jpg", "images/agibotworld/751/908652/cam_left_wrist/frame_000239.jpg", "images/agibotworld/751/908652/cam_right_wrist/frame_000239.jpg", "images/agibotworld/751/908652/cam_high/frame_000928.jpg", "images/agibotworld/751/908652/cam_left_wrist/frame_000928.jpg", "images/agibotworld/751/908652/cam_right_wrist/frame_000928.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Slice the noodles`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Slice the noodles`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 85.2, "hop_absolute_value": 85.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_563_897155-train-bf_000969-af_001037-plus-0000709", "task": "Remove the baked dessert from the oven.", "image": [ "images/agibotworld/563/901663/cam_high/frame_000020.jpg", "images/agibotworld/563/901663/cam_high/frame_002784.jpg", "images/agibotworld/563/897155/cam_high/frame_000969.jpg", "images/agibotworld/563/897155/cam_left_wrist/frame_000969.jpg", "images/agibotworld/563/897155/cam_right_wrist/frame_000969.jpg", "images/agibotworld/563/897155/cam_high/frame_001037.jpg", "images/agibotworld/563/897155/cam_left_wrist/frame_001037.jpg", "images/agibotworld/563/897155/cam_right_wrist/frame_001037.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the baked dessert from the oven.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the baked dessert from the oven.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.9, "hop_absolute_value": 1.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_580_869868-train-bf_007040-af_008647-plus-0000365", "task": "Make juice", "image": [ "images/agibotworld/580/869868/cam_high/frame_000047.jpg", "images/agibotworld/580/869868/cam_high/frame_011105.jpg", "images/agibotworld/580/869868/cam_high/frame_007040.jpg", "images/agibotworld/580/869868/cam_left_wrist/frame_007040.jpg", "images/agibotworld/580/869868/cam_right_wrist/frame_007040.jpg", "images/agibotworld/580/869868/cam_high/frame_008647.jpg", "images/agibotworld/580/869868/cam_left_wrist/frame_008647.jpg", "images/agibotworld/580/869868/cam_right_wrist/frame_008647.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make juice`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make juice`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 35.5, "hop_absolute_value": 35.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_714_863601-train-bf_002144-af_002243-plus-0000342", "task": "Open the door and turn off the light.", "image": [ "images/agibotworld/714/863601/cam_high/frame_000027.jpg", "images/agibotworld/714/863601/cam_high/frame_002440.jpg", "images/agibotworld/714/863601/cam_high/frame_002144.jpg", "images/agibotworld/714/863601/cam_left_wrist/frame_002144.jpg", "images/agibotworld/714/863601/cam_right_wrist/frame_002144.jpg", "images/agibotworld/714/863601/cam_high/frame_002243.jpg", "images/agibotworld/714/863601/cam_left_wrist/frame_002243.jpg", "images/agibotworld/714/863601/cam_right_wrist/frame_002243.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the door and turn off the light.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the door and turn off the light.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_537_877537-train-bf_000049-af_000318-plus-0000849", "task": "Make the bed", "image": [ "images/agibotworld/537/853970/cam_high/frame_000000.jpg", "images/agibotworld/537/853970/cam_high/frame_002854.jpg", "images/agibotworld/537/877537/cam_high/frame_000049.jpg", "images/agibotworld/537/877537/cam_left_wrist/frame_000049.jpg", "images/agibotworld/537/877537/cam_right_wrist/frame_000049.jpg", "images/agibotworld/537/877537/cam_high/frame_000318.jpg", "images/agibotworld/537/877537/cam_left_wrist/frame_000318.jpg", "images/agibotworld/537/877537/cam_right_wrist/frame_000318.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make the bed`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make the bed`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.9, "hop_absolute_value": 11.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_609_833098-train-bf_000885-af_001247-plus-0000355", "task": "Restock the hanging basket area", "image": [ "images/agibotworld/609/832144/cam_high/frame_000000.jpg", "images/agibotworld/609/832144/cam_high/frame_001100.jpg", "images/agibotworld/609/833098/cam_high/frame_000885.jpg", "images/agibotworld/609/833098/cam_left_wrist/frame_000885.jpg", "images/agibotworld/609/833098/cam_right_wrist/frame_000885.jpg", "images/agibotworld/609/833098/cam_high/frame_001247.jpg", "images/agibotworld/609/833098/cam_left_wrist/frame_001247.jpg", "images/agibotworld/609/833098/cam_right_wrist/frame_001247.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Restock the hanging basket area`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Restock the hanging basket area`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_444_716068-train-bf_001232-af_001368-plus-0000629", "task": "Fold short sleeve", "image": [ "images/agibotworld/444/716068/cam_high/frame_000000.jpg", "images/agibotworld/444/716068/cam_high/frame_004094.jpg", "images/agibotworld/444/716068/cam_high/frame_001232.jpg", "images/agibotworld/444/716068/cam_left_wrist/frame_001232.jpg", "images/agibotworld/444/716068/cam_right_wrist/frame_001232.jpg", "images/agibotworld/444/716068/cam_high/frame_001368.jpg", "images/agibotworld/444/716068/cam_left_wrist/frame_001368.jpg", "images/agibotworld/444/716068/cam_right_wrist/frame_001368.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold short sleeve`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold short sleeve`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_708_910844-train-bf_002674-af_002715-plus-0000627", "task": "Place name tags", "image": [ "images/agibotworld/708/919345/cam_high/frame_000019.jpg", "images/agibotworld/708/919345/cam_high/frame_002594.jpg", "images/agibotworld/708/910844/cam_high/frame_002674.jpg", "images/agibotworld/708/910844/cam_left_wrist/frame_002674.jpg", "images/agibotworld/708/910844/cam_right_wrist/frame_002674.jpg", "images/agibotworld/708/910844/cam_high/frame_002715.jpg", "images/agibotworld/708/910844/cam_left_wrist/frame_002715.jpg", "images/agibotworld/708/910844/cam_right_wrist/frame_002715.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place name tags`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place name tags`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.4, "hop_absolute_value": 2.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_445_793658-train-bf_000990-af_001318-plus-0000421", "task": "Open the fridge to get fruits and vegetables.", "image": [ "images/agibotworld/445/793658/cam_high/frame_000000.jpg", "images/agibotworld/445/793658/cam_high/frame_001755.jpg", "images/agibotworld/445/793658/cam_high/frame_000990.jpg", "images/agibotworld/445/793658/cam_left_wrist/frame_000990.jpg", "images/agibotworld/445/793658/cam_right_wrist/frame_000990.jpg", "images/agibotworld/445/793658/cam_high/frame_001318.jpg", "images/agibotworld/445/793658/cam_left_wrist/frame_001318.jpg", "images/agibotworld/445/793658/cam_right_wrist/frame_001318.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the fridge to get fruits and vegetables.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the fridge to get fruits and vegetables.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 63.2, "hop_absolute_value": 63.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_694_881816-train-bf_000596-af_002420-plus-0000562", "task": "Twist the bottle cap", "image": [ "images/agibotworld/694/882504/cam_high/frame_000000.jpg", "images/agibotworld/694/882504/cam_high/frame_002115.jpg", "images/agibotworld/694/881816/cam_high/frame_000596.jpg", "images/agibotworld/694/881816/cam_left_wrist/frame_000596.jpg", "images/agibotworld/694/881816/cam_right_wrist/frame_000596.jpg", "images/agibotworld/694/881816/cam_high/frame_002420.jpg", "images/agibotworld/694/881816/cam_left_wrist/frame_002420.jpg", "images/agibotworld/694/881816/cam_right_wrist/frame_002420.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_619_839706-train-bf_000362-af_000395-plus-0000096", "task": "Grab the toy", "image": [ "images/agibotworld/619/839476/cam_high/frame_000000.jpg", "images/agibotworld/619/839476/cam_high/frame_000409.jpg", "images/agibotworld/619/839706/cam_high/frame_000362.jpg", "images/agibotworld/619/839706/cam_left_wrist/frame_000362.jpg", "images/agibotworld/619/839706/cam_right_wrist/frame_000362.jpg", "images/agibotworld/619/839706/cam_high/frame_000395.jpg", "images/agibotworld/619/839706/cam_left_wrist/frame_000395.jpg", "images/agibotworld/619/839706/cam_right_wrist/frame_000395.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab the toy`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab the toy`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_428_744949-train-bf_000972-af_001327-plus-0000335", "task": "Open drawer and store items", "image": [ "images/agibotworld/428/733775/cam_high/frame_000000.jpg", "images/agibotworld/428/733775/cam_high/frame_002571.jpg", "images/agibotworld/428/744949/cam_high/frame_000972.jpg", "images/agibotworld/428/744949/cam_left_wrist/frame_000972.jpg", "images/agibotworld/428/744949/cam_right_wrist/frame_000972.jpg", "images/agibotworld/428/744949/cam_high/frame_001327.jpg", "images/agibotworld/428/744949/cam_left_wrist/frame_001327.jpg", "images/agibotworld/428/744949/cam_right_wrist/frame_001327.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open drawer and store items`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open drawer and store items`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_483_734313-train-bf_003157-af_003200-plus-0000937", "task": "Place items from the meeting room table into the storage box.", "image": [ "images/agibotworld/483/748025/cam_high/frame_000000.jpg", "images/agibotworld/483/748025/cam_high/frame_003381.jpg", "images/agibotworld/483/734313/cam_high/frame_003157.jpg", "images/agibotworld/483/734313/cam_left_wrist/frame_003157.jpg", "images/agibotworld/483/734313/cam_right_wrist/frame_003157.jpg", "images/agibotworld/483/734313/cam_high/frame_003200.jpg", "images/agibotworld/483/734313/cam_left_wrist/frame_003200.jpg", "images/agibotworld/483/734313/cam_right_wrist/frame_003200.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place items from the meeting room table into the storage box.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place items from the meeting room table into the storage box.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 7.1, "hop_absolute_value": 7.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_487_747514-train-bf_001430-af_001561-plus-0000313", "task": "Remove bottled water from the carton and arrange neatly on the table.", "image": [ "images/agibotworld/487/733038/cam_high/frame_000120.jpg", "images/agibotworld/487/733038/cam_high/frame_004704.jpg", "images/agibotworld/487/747514/cam_high/frame_001430.jpg", "images/agibotworld/487/747514/cam_left_wrist/frame_001430.jpg", "images/agibotworld/487/747514/cam_right_wrist/frame_001430.jpg", "images/agibotworld/487/747514/cam_high/frame_001561.jpg", "images/agibotworld/487/747514/cam_left_wrist/frame_001561.jpg", "images/agibotworld/487/747514/cam_right_wrist/frame_001561.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove bottled water from the carton and arrange neatly on the table.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove bottled water from the carton and arrange neatly on the table.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 6.4, "hop_absolute_value": 6.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_683_869069-train-bf_000774-af_001474-plus-0000942", "task": "Serve the meal", "image": [ "images/agibotworld/683/871173/cam_high/frame_000139.jpg", "images/agibotworld/683/871173/cam_high/frame_002427.jpg", "images/agibotworld/683/869069/cam_high/frame_000774.jpg", "images/agibotworld/683/869069/cam_left_wrist/frame_000774.jpg", "images/agibotworld/683/869069/cam_right_wrist/frame_000774.jpg", "images/agibotworld/683/869069/cam_high/frame_001474.jpg", "images/agibotworld/683/869069/cam_left_wrist/frame_001474.jpg", "images/agibotworld/683/869069/cam_right_wrist/frame_001474.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Serve the meal`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Serve the meal`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 43.8, "hop_absolute_value": 43.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_388_690998-train-bf_000788-af_000898-plus-0000730", "task": "Pickup in the supermarket", "image": [ "images/agibotworld/388/690721/cam_high/frame_000024.jpg", "images/agibotworld/388/690721/cam_high/frame_001585.jpg", "images/agibotworld/388/690998/cam_high/frame_000788.jpg", "images/agibotworld/388/690998/cam_left_wrist/frame_000788.jpg", "images/agibotworld/388/690998/cam_right_wrist/frame_000788.jpg", "images/agibotworld/388/690998/cam_high/frame_000898.jpg", "images/agibotworld/388/690998/cam_left_wrist/frame_000898.jpg", "images/agibotworld/388/690998/cam_right_wrist/frame_000898.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pickup in the supermarket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pickup in the supermarket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 13.3, "hop_absolute_value": 13.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_545_852733-train-bf_001436-af_001479-plus-0000783", "task": "Wash the dishes and silverware", "image": [ "images/agibotworld/545/854616/cam_high/frame_000000.jpg", "images/agibotworld/545/854616/cam_high/frame_003969.jpg", "images/agibotworld/545/852733/cam_high/frame_001436.jpg", "images/agibotworld/545/852733/cam_left_wrist/frame_001436.jpg", "images/agibotworld/545/852733/cam_right_wrist/frame_001436.jpg", "images/agibotworld/545/852733/cam_high/frame_001479.jpg", "images/agibotworld/545/852733/cam_left_wrist/frame_001479.jpg", "images/agibotworld/545/852733/cam_right_wrist/frame_001479.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wash the dishes and silverware`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wash the dishes and silverware`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.8, "hop_absolute_value": 1.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_378_670004-train-bf_000261-af_000318-plus-0000647", "task": "Clear table in the restaurant", "image": [ "images/agibotworld/378/670004/cam_high/frame_000000.jpg", "images/agibotworld/378/670004/cam_high/frame_001789.jpg", "images/agibotworld/378/670004/cam_high/frame_000261.jpg", "images/agibotworld/378/670004/cam_left_wrist/frame_000261.jpg", "images/agibotworld/378/670004/cam_right_wrist/frame_000261.jpg", "images/agibotworld/378/670004/cam_high/frame_000318.jpg", "images/agibotworld/378/670004/cam_left_wrist/frame_000318.jpg", "images/agibotworld/378/670004/cam_right_wrist/frame_000318.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Clear table in the restaurant`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Clear table in the restaurant`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.2, "hop_absolute_value": 3.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_365_692718-train-bf_000148-af_000727-plus-0000867", "task": "Sort personal care products", "image": [ "images/agibotworld/365/693581/cam_high/frame_000000.jpg", "images/agibotworld/365/693581/cam_high/frame_000882.jpg", "images/agibotworld/365/692718/cam_high/frame_000148.jpg", "images/agibotworld/365/692718/cam_left_wrist/frame_000148.jpg", "images/agibotworld/365/692718/cam_right_wrist/frame_000148.jpg", "images/agibotworld/365/692718/cam_high/frame_000727.jpg", "images/agibotworld/365/692718/cam_left_wrist/frame_000727.jpg", "images/agibotworld/365/692718/cam_right_wrist/frame_000727.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort personal care products`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort personal care products`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.4, "hop_absolute_value": 70.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_664_863238-train-bf_000190-af_000196-plus-0000836", "task": "Hammer the toy", "image": [ "images/agibotworld/664/866479/cam_high/frame_000021.jpg", "images/agibotworld/664/866479/cam_high/frame_000387.jpg", "images/agibotworld/664/863238/cam_high/frame_000190.jpg", "images/agibotworld/664/863238/cam_left_wrist/frame_000190.jpg", "images/agibotworld/664/863238/cam_right_wrist/frame_000190.jpg", "images/agibotworld/664/863238/cam_high/frame_000196.jpg", "images/agibotworld/664/863238/cam_left_wrist/frame_000196.jpg", "images/agibotworld/664/863238/cam_right_wrist/frame_000196.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Hammer the toy`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Hammer the toy`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.9, "hop_absolute_value": 5.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_433_726834-train-bf_000130-af_003525-plus-0000566", "task": "Make instant coffee", "image": [ "images/agibotworld/433/748897/cam_high/frame_000000.jpg", "images/agibotworld/433/748897/cam_high/frame_003362.jpg", "images/agibotworld/433/726834/cam_high/frame_000130.jpg", "images/agibotworld/433/726834/cam_left_wrist/frame_000130.jpg", "images/agibotworld/433/726834/cam_right_wrist/frame_000130.jpg", "images/agibotworld/433/726834/cam_high/frame_003525.jpg", "images/agibotworld/433/726834/cam_left_wrist/frame_003525.jpg", "images/agibotworld/433/726834/cam_right_wrist/frame_003525.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make instant coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make instant coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 90.1, "hop_absolute_value": 90.1, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_613_840809-train-bf_000000-af_000457-plus-0000670", "task": "Insert the pen cap", "image": [ "images/agibotworld/613/840809/cam_high/frame_000000.jpg", "images/agibotworld/613/840809/cam_high/frame_001136.jpg", "images/agibotworld/613/840809/cam_high/frame_000000.jpg", "images/agibotworld/613/840809/cam_left_wrist/frame_000000.jpg", "images/agibotworld/613/840809/cam_right_wrist/frame_000000.jpg", "images/agibotworld/613/840809/cam_high/frame_000457.jpg", "images/agibotworld/613/840809/cam_left_wrist/frame_000457.jpg", "images/agibotworld/613/840809/cam_right_wrist/frame_000457.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert the pen cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert the pen cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 26.7, "hop_absolute_value": 26.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_529_760498-train-bf_000116-af_000751-plus-0000307", "task": "Wipe the toilet with a cloth.", "image": [ "images/agibotworld/529/761293/cam_high/frame_000000.jpg", "images/agibotworld/529/761293/cam_high/frame_001496.jpg", "images/agibotworld/529/760498/cam_high/frame_000116.jpg", "images/agibotworld/529/760498/cam_left_wrist/frame_000116.jpg", "images/agibotworld/529/760498/cam_right_wrist/frame_000116.jpg", "images/agibotworld/529/760498/cam_high/frame_000751.jpg", "images/agibotworld/529/760498/cam_left_wrist/frame_000751.jpg", "images/agibotworld/529/760498/cam_right_wrist/frame_000751.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wipe the toilet with a cloth.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wipe the toilet with a cloth.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 47.8, "hop_absolute_value": 47.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_434_811045-train-bf_001425-af_001574-plus-0000754", "task": "Remove capsules from the coffee machine.", "image": [ "images/agibotworld/434/805262/cam_high/frame_000000.jpg", "images/agibotworld/434/805262/cam_high/frame_001539.jpg", "images/agibotworld/434/811045/cam_high/frame_001425.jpg", "images/agibotworld/434/811045/cam_left_wrist/frame_001425.jpg", "images/agibotworld/434/811045/cam_right_wrist/frame_001425.jpg", "images/agibotworld/434/811045/cam_high/frame_001574.jpg", "images/agibotworld/434/811045/cam_left_wrist/frame_001574.jpg", "images/agibotworld/434/811045/cam_right_wrist/frame_001574.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove capsules from the coffee machine.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove capsules from the coffee machine.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_367_649878-train-bf_000166-af_000393-plus-0000266", "task": "Take toast from toaster", "image": [ "images/agibotworld/367/649840/cam_high/frame_000013.jpg", "images/agibotworld/367/649840/cam_high/frame_000699.jpg", "images/agibotworld/367/649878/cam_high/frame_000166.jpg", "images/agibotworld/367/649878/cam_left_wrist/frame_000166.jpg", "images/agibotworld/367/649878/cam_right_wrist/frame_000166.jpg", "images/agibotworld/367/649878/cam_high/frame_000393.jpg", "images/agibotworld/367/649878/cam_left_wrist/frame_000393.jpg", "images/agibotworld/367/649878/cam_right_wrist/frame_000393.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take toast from toaster`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take toast from toaster`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 31.6, "hop_absolute_value": 31.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_544_819003-train-bf_000097-af_000821-plus-0000340", "task": "Wash the pot and spatula", "image": [ "images/agibotworld/544/819591/cam_high/frame_000000.jpg", "images/agibotworld/544/819591/cam_high/frame_003307.jpg", "images/agibotworld/544/819003/cam_high/frame_000097.jpg", "images/agibotworld/544/819003/cam_left_wrist/frame_000097.jpg", "images/agibotworld/544/819003/cam_right_wrist/frame_000097.jpg", "images/agibotworld/544/819003/cam_high/frame_000821.jpg", "images/agibotworld/544/819003/cam_left_wrist/frame_000821.jpg", "images/agibotworld/544/819003/cam_right_wrist/frame_000821.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wash the pot and spatula`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wash the pot and spatula`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 27.6, "hop_absolute_value": 27.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_445_805999-train-bf_000863-af_000940-plus-0000626", "task": "Open the fridge to get fruits and vegetables.", "image": [ "images/agibotworld/445/817496/cam_high/frame_000000.jpg", "images/agibotworld/445/817496/cam_high/frame_001784.jpg", "images/agibotworld/445/805999/cam_high/frame_000863.jpg", "images/agibotworld/445/805999/cam_left_wrist/frame_000863.jpg", "images/agibotworld/445/805999/cam_right_wrist/frame_000863.jpg", "images/agibotworld/445/805999/cam_high/frame_000940.jpg", "images/agibotworld/445/805999/cam_left_wrist/frame_000940.jpg", "images/agibotworld/445/805999/cam_right_wrist/frame_000940.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the fridge to get fruits and vegetables.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the fridge to get fruits and vegetables.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.5, "hop_absolute_value": 11.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_751_897259-train-bf_000483-af_000644-plus-0000311", "task": "Slice the noodles", "image": [ "images/agibotworld/751/897259/cam_high/frame_000034.jpg", "images/agibotworld/751/897259/cam_high/frame_001143.jpg", "images/agibotworld/751/897259/cam_high/frame_000483.jpg", "images/agibotworld/751/897259/cam_left_wrist/frame_000483.jpg", "images/agibotworld/751/897259/cam_right_wrist/frame_000483.jpg", "images/agibotworld/751/897259/cam_high/frame_000644.jpg", "images/agibotworld/751/897259/cam_left_wrist/frame_000644.jpg", "images/agibotworld/751/897259/cam_right_wrist/frame_000644.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Slice the noodles`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Slice the noodles`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 30.4, "hop_absolute_value": 30.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_555_787788-train-bf_000000-af_001371-plus-0000137", "task": "Fold shorts", "image": [ "images/agibotworld/555/801265/cam_high/frame_000000.jpg", "images/agibotworld/555/801265/cam_high/frame_001703.jpg", "images/agibotworld/555/787788/cam_high/frame_000000.jpg", "images/agibotworld/555/787788/cam_left_wrist/frame_000000.jpg", "images/agibotworld/555/787788/cam_right_wrist/frame_000000.jpg", "images/agibotworld/555/787788/cam_high/frame_001371.jpg", "images/agibotworld/555/787788/cam_left_wrist/frame_001371.jpg", "images/agibotworld/555/787788/cam_right_wrist/frame_001371.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_732_891682-train-bf_000546-af_001063-plus-0000286", "task": "Roll the dough", "image": [ "images/agibotworld/732/889077/cam_high/frame_000036.jpg", "images/agibotworld/732/889077/cam_high/frame_001215.jpg", "images/agibotworld/732/891682/cam_high/frame_000546.jpg", "images/agibotworld/732/891682/cam_left_wrist/frame_000546.jpg", "images/agibotworld/732/891682/cam_right_wrist/frame_000546.jpg", "images/agibotworld/732/891682/cam_high/frame_001063.jpg", "images/agibotworld/732/891682/cam_left_wrist/frame_001063.jpg", "images/agibotworld/732/891682/cam_right_wrist/frame_001063.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Roll the dough`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Roll the dough`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 38.9, "hop_absolute_value": 38.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_547_845046-train-bf_000117-af_001765-plus-0000423", "task": "Sort light and dark clothes", "image": [ "images/agibotworld/547/843117/cam_high/frame_000000.jpg", "images/agibotworld/547/843117/cam_high/frame_003161.jpg", "images/agibotworld/547/845046/cam_high/frame_000117.jpg", "images/agibotworld/547/845046/cam_left_wrist/frame_000117.jpg", "images/agibotworld/547/845046/cam_right_wrist/frame_000117.jpg", "images/agibotworld/547/845046/cam_high/frame_001765.jpg", "images/agibotworld/547/845046/cam_left_wrist/frame_001765.jpg", "images/agibotworld/547/845046/cam_right_wrist/frame_001765.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort light and dark clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort light and dark clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 56.2, "hop_absolute_value": 56.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_737_877746-train-bf_000694-af_000776-plus-0000086", "task": "Peel the skin", "image": [ "images/agibotworld/737/930010/cam_high/frame_000063.jpg", "images/agibotworld/737/930010/cam_high/frame_000887.jpg", "images/agibotworld/737/877746/cam_high/frame_000694.jpg", "images/agibotworld/737/877746/cam_left_wrist/frame_000694.jpg", "images/agibotworld/737/877746/cam_right_wrist/frame_000694.jpg", "images/agibotworld/737/877746/cam_high/frame_000776.jpg", "images/agibotworld/737/877746/cam_left_wrist/frame_000776.jpg", "images/agibotworld/737/877746/cam_right_wrist/frame_000776.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Peel the skin`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Peel the skin`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_369_799945-train-bf_001148-af_001444-plus-0000423", "task": "Remove clothes from the washing machine", "image": [ "images/agibotworld/369/799945/cam_high/frame_000000.jpg", "images/agibotworld/369/799945/cam_high/frame_001838.jpg", "images/agibotworld/369/799945/cam_high/frame_001148.jpg", "images/agibotworld/369/799945/cam_left_wrist/frame_001148.jpg", "images/agibotworld/369/799945/cam_right_wrist/frame_001148.jpg", "images/agibotworld/369/799945/cam_high/frame_001444.jpg", "images/agibotworld/369/799945/cam_left_wrist/frame_001444.jpg", "images/agibotworld/369/799945/cam_right_wrist/frame_001444.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove clothes from the washing machine`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove clothes from the washing machine`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_681_857678-train-bf_000000-af_001377-plus-0000137", "task": "Fold the shorts", "image": [ "images/agibotworld/681/858642/cam_high/frame_000000.jpg", "images/agibotworld/681/858642/cam_high/frame_001296.jpg", "images/agibotworld/681/857678/cam_high/frame_000000.jpg", "images/agibotworld/681/857678/cam_left_wrist/frame_000000.jpg", "images/agibotworld/681/857678/cam_right_wrist/frame_000000.jpg", "images/agibotworld/681/857678/cam_high/frame_001377.jpg", "images/agibotworld/681/857678/cam_left_wrist/frame_001377.jpg", "images/agibotworld/681/857678/cam_right_wrist/frame_001377.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_480_723559-train-bf_002638-af_003090-plus-0000531", "task": "Make sandwiches with salad dressing 1213 edition", "image": [ "images/agibotworld/480/710283/cam_high/frame_000023.jpg", "images/agibotworld/480/710283/cam_high/frame_002265.jpg", "images/agibotworld/480/723559/cam_high/frame_002638.jpg", "images/agibotworld/480/723559/cam_left_wrist/frame_002638.jpg", "images/agibotworld/480/723559/cam_right_wrist/frame_002638.jpg", "images/agibotworld/480/723559/cam_high/frame_003090.jpg", "images/agibotworld/480/723559/cam_left_wrist/frame_003090.jpg", "images/agibotworld/480/723559/cam_right_wrist/frame_003090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make sandwiches with salad dressing 1213 edition`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make sandwiches with salad dressing 1213 edition`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_570_794302-train-bf_000462-af_001760-plus-0000310", "task": "Fold the T-shirt on the field", "image": [ "images/agibotworld/570/794302/cam_high/frame_000000.jpg", "images/agibotworld/570/794302/cam_high/frame_001760.jpg", "images/agibotworld/570/794302/cam_high/frame_000462.jpg", "images/agibotworld/570/794302/cam_left_wrist/frame_000462.jpg", "images/agibotworld/570/794302/cam_right_wrist/frame_000462.jpg", "images/agibotworld/570/794302/cam_high/frame_001760.jpg", "images/agibotworld/570/794302/cam_left_wrist/frame_001760.jpg", "images/agibotworld/570/794302/cam_right_wrist/frame_001760.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the T-shirt on the field`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the T-shirt on the field`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_666_867023-train-bf_000000-af_000554-plus-0000336", "task": "Close the pen cap", "image": [ "images/agibotworld/666/876500/cam_high/frame_000051.jpg", "images/agibotworld/666/876500/cam_high/frame_001161.jpg", "images/agibotworld/666/867023/cam_high/frame_000000.jpg", "images/agibotworld/666/867023/cam_left_wrist/frame_000000.jpg", "images/agibotworld/666/867023/cam_right_wrist/frame_000000.jpg", "images/agibotworld/666/867023/cam_high/frame_000554.jpg", "images/agibotworld/666/867023/cam_left_wrist/frame_000554.jpg", "images/agibotworld/666/867023/cam_right_wrist/frame_000554.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the pen cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the pen cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 43.3, "hop_absolute_value": 43.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_494_717290-train-bf_000686-af_002312-plus-0000514", "task": "Organize the kitchen counter seasonings", "image": [ "images/agibotworld/494/722427/cam_high/frame_000000.jpg", "images/agibotworld/494/722427/cam_high/frame_002816.jpg", "images/agibotworld/494/717290/cam_high/frame_000686.jpg", "images/agibotworld/494/717290/cam_left_wrist/frame_000686.jpg", "images/agibotworld/494/717290/cam_right_wrist/frame_000686.jpg", "images/agibotworld/494/717290/cam_high/frame_002312.jpg", "images/agibotworld/494/717290/cam_left_wrist/frame_002312.jpg", "images/agibotworld/494/717290/cam_right_wrist/frame_002312.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Organize the kitchen counter seasonings`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Organize the kitchen counter seasonings`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 81.4, "hop_absolute_value": 81.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_609_833536-train-bf_000170-af_000693-plus-0000337", "task": "Restock the hanging basket area", "image": [ "images/agibotworld/609/833536/cam_high/frame_000000.jpg", "images/agibotworld/609/833536/cam_high/frame_001022.jpg", "images/agibotworld/609/833536/cam_high/frame_000170.jpg", "images/agibotworld/609/833536/cam_left_wrist/frame_000170.jpg", "images/agibotworld/609/833536/cam_right_wrist/frame_000170.jpg", "images/agibotworld/609/833536/cam_high/frame_000693.jpg", "images/agibotworld/609/833536/cam_left_wrist/frame_000693.jpg", "images/agibotworld/609/833536/cam_right_wrist/frame_000693.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Restock the hanging basket area`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Restock the hanging basket area`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_445_825190-train-bf_000894-af_001912-plus-0000515", "task": "Open the fridge to get fruits and vegetables.", "image": [ "images/agibotworld/445/785448/cam_high/frame_000031.jpg", "images/agibotworld/445/785448/cam_high/frame_002206.jpg", "images/agibotworld/445/825190/cam_high/frame_000894.jpg", "images/agibotworld/445/825190/cam_left_wrist/frame_000894.jpg", "images/agibotworld/445/825190/cam_right_wrist/frame_000894.jpg", "images/agibotworld/445/825190/cam_high/frame_001912.jpg", "images/agibotworld/445/825190/cam_left_wrist/frame_001912.jpg", "images/agibotworld/445/825190/cam_right_wrist/frame_001912.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the fridge to get fruits and vegetables.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the fridge to get fruits and vegetables.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 88.5, "hop_absolute_value": 88.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_414_673851-train-bf_001845-af_002761-plus-0000482", "task": "Hang clothes with hanger", "image": [ "images/agibotworld/414/710360/cam_high/frame_000028.jpg", "images/agibotworld/414/710360/cam_high/frame_001928.jpg", "images/agibotworld/414/673851/cam_high/frame_001845.jpg", "images/agibotworld/414/673851/cam_left_wrist/frame_001845.jpg", "images/agibotworld/414/673851/cam_right_wrist/frame_001845.jpg", "images/agibotworld/414/673851/cam_high/frame_002761.jpg", "images/agibotworld/414/673851/cam_left_wrist/frame_002761.jpg", "images/agibotworld/414/673851/cam_right_wrist/frame_002761.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Hang clothes with hanger`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Hang clothes with hanger`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_466_826388-train-bf_001180-af_002023-plus-0000564", "task": "Separate dark and light clothes", "image": [ "images/agibotworld/466/826388/cam_high/frame_000000.jpg", "images/agibotworld/466/826388/cam_high/frame_002088.jpg", "images/agibotworld/466/826388/cam_high/frame_001180.jpg", "images/agibotworld/466/826388/cam_left_wrist/frame_001180.jpg", "images/agibotworld/466/826388/cam_right_wrist/frame_001180.jpg", "images/agibotworld/466/826388/cam_high/frame_002023.jpg", "images/agibotworld/466/826388/cam_left_wrist/frame_002023.jpg", "images/agibotworld/466/826388/cam_right_wrist/frame_002023.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Separate dark and light clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Separate dark and light clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 93.3, "hop_absolute_value": 93.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_485_714626-train-bf_000388-af_000421-plus-0000279", "task": "Pick up the scrap paper on the desk and feed it into the shredder for destruction.", "image": [ "images/agibotworld/485/716819/cam_high/frame_000000.jpg", "images/agibotworld/485/716819/cam_high/frame_002226.jpg", "images/agibotworld/485/714626/cam_high/frame_000388.jpg", "images/agibotworld/485/714626/cam_left_wrist/frame_000388.jpg", "images/agibotworld/485/714626/cam_right_wrist/frame_000388.jpg", "images/agibotworld/485/714626/cam_high/frame_000421.jpg", "images/agibotworld/485/714626/cam_left_wrist/frame_000421.jpg", "images/agibotworld/485/714626/cam_right_wrist/frame_000421.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the scrap paper on the desk and feed it into the shredder for destruction.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the scrap paper on the desk and feed it into the shredder for destruction.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.8, "hop_absolute_value": 4.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_385_680599-train-bf_000590-af_000864-plus-0000390", "task": "Pickup in the supermarket produce section", "image": [ "images/agibotworld/385/680599/cam_high/frame_000000.jpg", "images/agibotworld/385/680599/cam_high/frame_001119.jpg", "images/agibotworld/385/680599/cam_high/frame_000590.jpg", "images/agibotworld/385/680599/cam_left_wrist/frame_000590.jpg", "images/agibotworld/385/680599/cam_right_wrist/frame_000590.jpg", "images/agibotworld/385/680599/cam_high/frame_000864.jpg", "images/agibotworld/385/680599/cam_left_wrist/frame_000864.jpg", "images/agibotworld/385/680599/cam_right_wrist/frame_000864.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pickup in the supermarket produce section`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pickup in the supermarket produce section`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.9, "hop_absolute_value": 57.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_434_824558-train-bf_000555-af_000591-plus-0000567", "task": "Remove capsules from the coffee machine.", "image": [ "images/agibotworld/434/824558/cam_high/frame_000000.jpg", "images/agibotworld/434/824558/cam_high/frame_001356.jpg", "images/agibotworld/434/824558/cam_high/frame_000555.jpg", "images/agibotworld/434/824558/cam_left_wrist/frame_000555.jpg", "images/agibotworld/434/824558/cam_right_wrist/frame_000555.jpg", "images/agibotworld/434/824558/cam_high/frame_000591.jpg", "images/agibotworld/434/824558/cam_left_wrist/frame_000591.jpg", "images/agibotworld/434/824558/cam_right_wrist/frame_000591.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove capsules from the coffee machine.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove capsules from the coffee machine.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.9, "hop_absolute_value": 5.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_471_712816-train-bf_000456-af_000499-plus-0000624", "task": "Organize the condiments on the stove.", "image": [ "images/agibotworld/471/712816/cam_high/frame_000000.jpg", "images/agibotworld/471/712816/cam_high/frame_003464.jpg", "images/agibotworld/471/712816/cam_high/frame_000456.jpg", "images/agibotworld/471/712816/cam_left_wrist/frame_000456.jpg", "images/agibotworld/471/712816/cam_right_wrist/frame_000456.jpg", "images/agibotworld/471/712816/cam_high/frame_000499.jpg", "images/agibotworld/471/712816/cam_left_wrist/frame_000499.jpg", "images/agibotworld/471/712816/cam_right_wrist/frame_000499.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Organize the condiments on the stove.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Organize the condiments on the stove.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.0, "hop_absolute_value": 2.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_551_870200-train-bf_000394-af_000458-plus-0000600", "task": "Disinfect the countertop", "image": [ "images/agibotworld/551/865375/cam_high/frame_000000.jpg", "images/agibotworld/551/865375/cam_high/frame_001981.jpg", "images/agibotworld/551/870200/cam_high/frame_000394.jpg", "images/agibotworld/551/870200/cam_left_wrist/frame_000394.jpg", "images/agibotworld/551/870200/cam_right_wrist/frame_000394.jpg", "images/agibotworld/551/870200/cam_high/frame_000458.jpg", "images/agibotworld/551/870200/cam_left_wrist/frame_000458.jpg", "images/agibotworld/551/870200/cam_right_wrist/frame_000458.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disinfect the countertop`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disinfect the countertop`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.1, "hop_absolute_value": 3.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_726_920965-train-bf_000383-af_000410-plus-0000635", "task": "Lift dumbbells", "image": [ "images/agibotworld/726/887675/cam_high/frame_000000.jpg", "images/agibotworld/726/887675/cam_high/frame_000904.jpg", "images/agibotworld/726/920965/cam_high/frame_000383.jpg", "images/agibotworld/726/920965/cam_left_wrist/frame_000383.jpg", "images/agibotworld/726/920965/cam_right_wrist/frame_000383.jpg", "images/agibotworld/726/920965/cam_high/frame_000410.jpg", "images/agibotworld/726/920965/cam_left_wrist/frame_000410.jpg", "images/agibotworld/726/920965/cam_right_wrist/frame_000410.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift dumbbells`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift dumbbells`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.5, "hop_absolute_value": 4.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_570_837778-train-bf_001456-af_001683-plus-0000188", "task": "Fold the T-shirt on the field", "image": [ "images/agibotworld/570/821568/cam_high/frame_000000.jpg", "images/agibotworld/570/821568/cam_high/frame_002631.jpg", "images/agibotworld/570/837778/cam_high/frame_001456.jpg", "images/agibotworld/570/837778/cam_left_wrist/frame_001456.jpg", "images/agibotworld/570/837778/cam_right_wrist/frame_001456.jpg", "images/agibotworld/570/837778/cam_high/frame_001683.jpg", "images/agibotworld/570/837778/cam_left_wrist/frame_001683.jpg", "images/agibotworld/570/837778/cam_right_wrist/frame_001683.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the T-shirt on the field`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the T-shirt on the field`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_451_830616-train-bf_000981-af_001006-plus-0000922", "task": "Make a salad", "image": [ "images/agibotworld/451/810376/cam_high/frame_000000.jpg", "images/agibotworld/451/810376/cam_high/frame_004044.jpg", "images/agibotworld/451/830616/cam_high/frame_000981.jpg", "images/agibotworld/451/830616/cam_left_wrist/frame_000981.jpg", "images/agibotworld/451/830616/cam_right_wrist/frame_000981.jpg", "images/agibotworld/451/830616/cam_high/frame_001006.jpg", "images/agibotworld/451/830616/cam_left_wrist/frame_001006.jpg", "images/agibotworld/451/830616/cam_right_wrist/frame_001006.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make a salad`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make a salad`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.0, "hop_absolute_value": 1.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_444_713454-train-bf_000261-af_000340-plus-0000268", "task": "Fold short sleeve", "image": [ "images/agibotworld/444/714373/cam_high/frame_000000.jpg", "images/agibotworld/444/714373/cam_high/frame_004360.jpg", "images/agibotworld/444/713454/cam_high/frame_000261.jpg", "images/agibotworld/444/713454/cam_left_wrist/frame_000261.jpg", "images/agibotworld/444/713454/cam_right_wrist/frame_000261.jpg", "images/agibotworld/444/713454/cam_high/frame_000340.jpg", "images/agibotworld/444/713454/cam_left_wrist/frame_000340.jpg", "images/agibotworld/444/713454/cam_right_wrist/frame_000340.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold short sleeve`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold short sleeve`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.7, "hop_absolute_value": 3.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_712_875210-train-bf_000531-af_000572-plus-0000616", "task": "Carry bottled water.", "image": [ "images/agibotworld/712/877362/cam_high/frame_000000.jpg", "images/agibotworld/712/877362/cam_high/frame_001771.jpg", "images/agibotworld/712/875210/cam_high/frame_000531.jpg", "images/agibotworld/712/875210/cam_left_wrist/frame_000531.jpg", "images/agibotworld/712/875210/cam_right_wrist/frame_000531.jpg", "images/agibotworld/712/875210/cam_high/frame_000572.jpg", "images/agibotworld/712/875210/cam_left_wrist/frame_000572.jpg", "images/agibotworld/712/875210/cam_right_wrist/frame_000572.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Carry bottled water.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Carry bottled water.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.4, "hop_absolute_value": 3.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_434_823953-train-bf_000071-af_000470-plus-0000682", "task": "Remove capsules from the coffee machine.", "image": [ "images/agibotworld/434/830816/cam_high/frame_000000.jpg", "images/agibotworld/434/830816/cam_high/frame_001379.jpg", "images/agibotworld/434/823953/cam_high/frame_000071.jpg", "images/agibotworld/434/823953/cam_left_wrist/frame_000071.jpg", "images/agibotworld/434/823953/cam_right_wrist/frame_000071.jpg", "images/agibotworld/434/823953/cam_high/frame_000470.jpg", "images/agibotworld/434/823953/cam_left_wrist/frame_000470.jpg", "images/agibotworld/434/823953/cam_right_wrist/frame_000470.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove capsules from the coffee machine.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove capsules from the coffee machine.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 27.6, "hop_absolute_value": 27.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_385_680737-train-bf_000679-af_000894-plus-0000931", "task": "Pickup in the supermarket produce section", "image": [ "images/agibotworld/385/680542/cam_high/frame_000000.jpg", "images/agibotworld/385/680542/cam_high/frame_001203.jpg", "images/agibotworld/385/680737/cam_high/frame_000679.jpg", "images/agibotworld/385/680737/cam_left_wrist/frame_000679.jpg", "images/agibotworld/385/680737/cam_right_wrist/frame_000679.jpg", "images/agibotworld/385/680737/cam_high/frame_000894.jpg", "images/agibotworld/385/680737/cam_left_wrist/frame_000894.jpg", "images/agibotworld/385/680737/cam_right_wrist/frame_000894.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pickup in the supermarket produce section`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pickup in the supermarket produce section`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 43.8, "hop_absolute_value": 43.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_664_857953-train-bf_000376-af_000442-plus-0000834", "task": "Hammer the toy", "image": [ "images/agibotworld/664/866782/cam_high/frame_000024.jpg", "images/agibotworld/664/866782/cam_high/frame_000359.jpg", "images/agibotworld/664/857953/cam_high/frame_000376.jpg", "images/agibotworld/664/857953/cam_left_wrist/frame_000376.jpg", "images/agibotworld/664/857953/cam_right_wrist/frame_000376.jpg", "images/agibotworld/664/857953/cam_high/frame_000442.jpg", "images/agibotworld/664/857953/cam_left_wrist/frame_000442.jpg", "images/agibotworld/664/857953/cam_right_wrist/frame_000442.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Hammer the toy`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Hammer the toy`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.8, "hop_absolute_value": 9.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_739_879278-train-bf_000155-af_000180-plus-0000162", "task": "Hand the menu", "image": [ "images/agibotworld/739/880599/cam_high/frame_000056.jpg", "images/agibotworld/739/880599/cam_high/frame_000512.jpg", "images/agibotworld/739/879278/cam_high/frame_000155.jpg", "images/agibotworld/739/879278/cam_left_wrist/frame_000155.jpg", "images/agibotworld/739/879278/cam_right_wrist/frame_000155.jpg", "images/agibotworld/739/879278/cam_high/frame_000180.jpg", "images/agibotworld/739/879278/cam_left_wrist/frame_000180.jpg", "images/agibotworld/739/879278/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Hand the menu`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Hand the menu`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 7.7, "hop_absolute_value": 7.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_599_837897-train-bf_000090-af_000181-plus-0000079", "task": "Fold the shorts", "image": [ "images/agibotworld/599/837897/cam_high/frame_000000.jpg", "images/agibotworld/599/837897/cam_high/frame_001156.jpg", "images/agibotworld/599/837897/cam_high/frame_000090.jpg", "images/agibotworld/599/837897/cam_left_wrist/frame_000090.jpg", "images/agibotworld/599/837897/cam_right_wrist/frame_000090.jpg", "images/agibotworld/599/837897/cam_high/frame_000181.jpg", "images/agibotworld/599/837897/cam_left_wrist/frame_000181.jpg", "images/agibotworld/599/837897/cam_right_wrist/frame_000181.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_434_831844-train-bf_000899-af_000164-minus-0000888", "task": "Remove capsules from the coffee machine.", "image": [ "images/agibotworld/434/831844/cam_high/frame_000000.jpg", "images/agibotworld/434/831844/cam_high/frame_001496.jpg", "images/agibotworld/434/831844/cam_high/frame_000899.jpg", "images/agibotworld/434/831844/cam_left_wrist/frame_000899.jpg", "images/agibotworld/434/831844/cam_right_wrist/frame_000899.jpg", "images/agibotworld/434/831844/cam_high/frame_000164.jpg", "images/agibotworld/434/831844/cam_left_wrist/frame_000164.jpg", "images/agibotworld/434/831844/cam_right_wrist/frame_000164.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove capsules from the coffee machine.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove capsules from the coffee machine.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -90.0, "hop_absolute_value": 90.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_607_834351-train-bf_001737-af_000458-minus-0000142", "task": "Scan and package the goods", "image": [ "images/agibotworld/607/834735/cam_high/frame_000000.jpg", "images/agibotworld/607/834735/cam_high/frame_001912.jpg", "images/agibotworld/607/834351/cam_high/frame_001737.jpg", "images/agibotworld/607/834351/cam_left_wrist/frame_001737.jpg", "images/agibotworld/607/834351/cam_right_wrist/frame_001737.jpg", "images/agibotworld/607/834351/cam_high/frame_000458.jpg", "images/agibotworld/607/834351/cam_left_wrist/frame_000458.jpg", "images/agibotworld/607/834351/cam_right_wrist/frame_000458.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Scan and package the goods`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Scan and package the goods`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -71.0, "hop_absolute_value": 71.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_548_834029-train-bf_001224-af_000342-minus-0000134", "task": "Separate dark and light clothes", "image": [ "images/agibotworld/548/835663/cam_high/frame_000000.jpg", "images/agibotworld/548/835663/cam_high/frame_001782.jpg", "images/agibotworld/548/834029/cam_high/frame_001224.jpg", "images/agibotworld/548/834029/cam_left_wrist/frame_001224.jpg", "images/agibotworld/548/834029/cam_right_wrist/frame_001224.jpg", "images/agibotworld/548/834029/cam_high/frame_000342.jpg", "images/agibotworld/548/834029/cam_left_wrist/frame_000342.jpg", "images/agibotworld/548/834029/cam_right_wrist/frame_000342.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Separate dark and light clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Separate dark and light clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_392_656821-train-bf_002467-af_002399-minus-0000833", "task": "Brush water bottle", "image": [ "images/agibotworld/392/693191/cam_high/frame_000042.jpg", "images/agibotworld/392/693191/cam_high/frame_002692.jpg", "images/agibotworld/392/656821/cam_high/frame_002467.jpg", "images/agibotworld/392/656821/cam_left_wrist/frame_002467.jpg", "images/agibotworld/392/656821/cam_right_wrist/frame_002467.jpg", "images/agibotworld/392/656821/cam_high/frame_002399.jpg", "images/agibotworld/392/656821/cam_left_wrist/frame_002399.jpg", "images/agibotworld/392/656821/cam_right_wrist/frame_002399.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Brush water bottle`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Brush water bottle`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.9, "hop_absolute_value": 1.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_542_775475-train-bf_001407-af_001320-minus-0000631", "task": "Boil water with a kettle", "image": [ "images/agibotworld/542/923437/cam_high/frame_000000.jpg", "images/agibotworld/542/923437/cam_high/frame_001686.jpg", "images/agibotworld/542/775475/cam_high/frame_001407.jpg", "images/agibotworld/542/775475/cam_left_wrist/frame_001407.jpg", "images/agibotworld/542/775475/cam_right_wrist/frame_001407.jpg", "images/agibotworld/542/775475/cam_high/frame_001320.jpg", "images/agibotworld/542/775475/cam_left_wrist/frame_001320.jpg", "images/agibotworld/542/775475/cam_right_wrist/frame_001320.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Boil water with a kettle`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Boil water with a kettle`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.2, "hop_absolute_value": 6.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_602_824201-train-bf_000456-af_000314-minus-0000186", "task": "Restock the hanging basket area", "image": [ "images/agibotworld/602/824201/cam_high/frame_000000.jpg", "images/agibotworld/602/824201/cam_high/frame_001163.jpg", "images/agibotworld/602/824201/cam_high/frame_000456.jpg", "images/agibotworld/602/824201/cam_left_wrist/frame_000456.jpg", "images/agibotworld/602/824201/cam_right_wrist/frame_000456.jpg", "images/agibotworld/602/824201/cam_high/frame_000314.jpg", "images/agibotworld/602/824201/cam_left_wrist/frame_000314.jpg", "images/agibotworld/602/824201/cam_right_wrist/frame_000314.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Restock the hanging basket area`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Restock the hanging basket area`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_474_768624-train-bf_001791-af_001165-minus-0000243", "task": "Arrange flowers", "image": [ "images/agibotworld/474/757975/cam_high/frame_000041.jpg", "images/agibotworld/474/757975/cam_high/frame_002326.jpg", "images/agibotworld/474/768624/cam_high/frame_001791.jpg", "images/agibotworld/474/768624/cam_left_wrist/frame_001791.jpg", "images/agibotworld/474/768624/cam_right_wrist/frame_001791.jpg", "images/agibotworld/474/768624/cam_high/frame_001165.jpg", "images/agibotworld/474/768624/cam_left_wrist/frame_001165.jpg", "images/agibotworld/474/768624/cam_right_wrist/frame_001165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Arrange flowers`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Arrange flowers`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -34.4, "hop_absolute_value": 34.4, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_462_855409-train-bf_001797-af_000977-minus-0000916", "task": "Discard the trash bag into the large bin.", "image": [ "images/agibotworld/462/837678/cam_high/frame_000000.jpg", "images/agibotworld/462/837678/cam_high/frame_002020.jpg", "images/agibotworld/462/855409/cam_high/frame_001797.jpg", "images/agibotworld/462/855409/cam_left_wrist/frame_001797.jpg", "images/agibotworld/462/855409/cam_right_wrist/frame_001797.jpg", "images/agibotworld/462/855409/cam_high/frame_000977.jpg", "images/agibotworld/462/855409/cam_left_wrist/frame_000977.jpg", "images/agibotworld/462/855409/cam_right_wrist/frame_000977.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Discard the trash bag into the large bin.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Discard the trash bag into the large bin.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -36.1, "hop_absolute_value": 36.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_372_672995-train-bf_000911-af_000751-minus-0000948", "task": "Packing in the supermarket", "image": [ "images/agibotworld/372/672833/cam_high/frame_000000.jpg", "images/agibotworld/372/672833/cam_high/frame_001866.jpg", "images/agibotworld/372/672995/cam_high/frame_000911.jpg", "images/agibotworld/372/672995/cam_left_wrist/frame_000911.jpg", "images/agibotworld/372/672995/cam_right_wrist/frame_000911.jpg", "images/agibotworld/372/672995/cam_high/frame_000751.jpg", "images/agibotworld/372/672995/cam_left_wrist/frame_000751.jpg", "images/agibotworld/372/672995/cam_right_wrist/frame_000751.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Packing in the supermarket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Packing in the supermarket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -18.5, "hop_absolute_value": 18.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_368_677756-train-bf_002172-af_002048-minus-0000766", "task": "Cook vegetables with oven", "image": [ "images/agibotworld/368/671476/cam_high/frame_000000.jpg", "images/agibotworld/368/671476/cam_high/frame_003224.jpg", "images/agibotworld/368/677756/cam_high/frame_002172.jpg", "images/agibotworld/368/677756/cam_left_wrist/frame_002172.jpg", "images/agibotworld/368/677756/cam_right_wrist/frame_002172.jpg", "images/agibotworld/368/677756/cam_high/frame_002048.jpg", "images/agibotworld/368/677756/cam_left_wrist/frame_002048.jpg", "images/agibotworld/368/677756/cam_right_wrist/frame_002048.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Cook vegetables with oven`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Cook vegetables with oven`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -8.2, "hop_absolute_value": 8.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_506_751110-train-bf_000763-af_000439-minus-0000160", "task": "Restock supermarket snacks", "image": [ "images/agibotworld/506/769642/cam_high/frame_000000.jpg", "images/agibotworld/506/769642/cam_high/frame_001268.jpg", "images/agibotworld/506/751110/cam_high/frame_000763.jpg", "images/agibotworld/506/751110/cam_left_wrist/frame_000763.jpg", "images/agibotworld/506/751110/cam_right_wrist/frame_000763.jpg", "images/agibotworld/506/751110/cam_high/frame_000439.jpg", "images/agibotworld/506/751110/cam_left_wrist/frame_000439.jpg", "images/agibotworld/506/751110/cam_right_wrist/frame_000439.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Restock supermarket snacks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Restock supermarket snacks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -44.4, "hop_absolute_value": 44.4, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_428_744118-train-bf_001736-af_001443-minus-0000802", "task": "Open drawer and store items", "image": [ "images/agibotworld/428/744118/cam_high/frame_000000.jpg", "images/agibotworld/428/744118/cam_high/frame_001998.jpg", "images/agibotworld/428/744118/cam_high/frame_001736.jpg", "images/agibotworld/428/744118/cam_left_wrist/frame_001736.jpg", "images/agibotworld/428/744118/cam_right_wrist/frame_001736.jpg", "images/agibotworld/428/744118/cam_high/frame_001443.jpg", "images/agibotworld/428/744118/cam_left_wrist/frame_001443.jpg", "images/agibotworld/428/744118/cam_right_wrist/frame_001443.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open drawer and store items`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open drawer and store items`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.3, "hop_absolute_value": 16.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_573_795941-train-bf_000716-af_000618-minus-0000905", "task": "Open the drawer to store items.", "image": [ "images/agibotworld/573/795941/cam_high/frame_000000.jpg", "images/agibotworld/573/795941/cam_high/frame_001376.jpg", "images/agibotworld/573/795941/cam_high/frame_000716.jpg", "images/agibotworld/573/795941/cam_left_wrist/frame_000716.jpg", "images/agibotworld/573/795941/cam_right_wrist/frame_000716.jpg", "images/agibotworld/573/795941/cam_high/frame_000618.jpg", "images/agibotworld/573/795941/cam_left_wrist/frame_000618.jpg", "images/agibotworld/573/795941/cam_right_wrist/frame_000618.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the drawer to store items.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the drawer to store items.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -41.7, "hop_absolute_value": 41.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_707_878753-train-bf_001064-af_000579-minus-0000188", "task": "Twist the bottle cap", "image": [ "images/agibotworld/707/881218/cam_high/frame_000000.jpg", "images/agibotworld/707/881218/cam_high/frame_001978.jpg", "images/agibotworld/707/878753/cam_high/frame_001064.jpg", "images/agibotworld/707/878753/cam_left_wrist/frame_001064.jpg", "images/agibotworld/707/878753/cam_right_wrist/frame_001064.jpg", "images/agibotworld/707/878753/cam_high/frame_000579.jpg", "images/agibotworld/707/878753/cam_left_wrist/frame_000579.jpg", "images/agibotworld/707/878753/cam_right_wrist/frame_000579.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -44.4, "hop_absolute_value": 44.4, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_446_690977-train-bf_000530-af_000000-minus-0000017", "task": "Transport table with another robot", "image": [ "images/agibotworld/446/687412/cam_high/frame_000000.jpg", "images/agibotworld/446/687412/cam_high/frame_000832.jpg", "images/agibotworld/446/690977/cam_high/frame_000530.jpg", "images/agibotworld/446/690977/cam_left_wrist/frame_000530.jpg", "images/agibotworld/446/690977/cam_right_wrist/frame_000530.jpg", "images/agibotworld/446/690977/cam_high/frame_000000.jpg", "images/agibotworld/446/690977/cam_left_wrist/frame_000000.jpg", "images/agibotworld/446/690977/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Transport table with another robot`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Transport table with another robot`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_480_719869-train-bf_002056-af_000692-minus-0000160", "task": "Make sandwiches with salad dressing 1213 edition", "image": [ "images/agibotworld/480/719869/cam_high/frame_000082.jpg", "images/agibotworld/480/719869/cam_high/frame_002625.jpg", "images/agibotworld/480/719869/cam_high/frame_002056.jpg", "images/agibotworld/480/719869/cam_left_wrist/frame_002056.jpg", "images/agibotworld/480/719869/cam_right_wrist/frame_002056.jpg", "images/agibotworld/480/719869/cam_high/frame_000692.jpg", "images/agibotworld/480/719869/cam_left_wrist/frame_000692.jpg", "images/agibotworld/480/719869/cam_right_wrist/frame_000692.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make sandwiches with salad dressing 1213 edition`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make sandwiches with salad dressing 1213 edition`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -61.2, "hop_absolute_value": 61.2, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_709_917512-train-bf_002352-af_002087-minus-0000931", "task": "Pack the medicines", "image": [ "images/agibotworld/709/891234/cam_high/frame_000000.jpg", "images/agibotworld/709/891234/cam_high/frame_004382.jpg", "images/agibotworld/709/917512/cam_high/frame_002352.jpg", "images/agibotworld/709/917512/cam_left_wrist/frame_002352.jpg", "images/agibotworld/709/917512/cam_right_wrist/frame_002352.jpg", "images/agibotworld/709/917512/cam_high/frame_002087.jpg", "images/agibotworld/709/917512/cam_left_wrist/frame_002087.jpg", "images/agibotworld/709/917512/cam_right_wrist/frame_002087.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pack the medicines`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pack the medicines`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -8.6, "hop_absolute_value": 8.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_455_715725-train-bf_000226-af_000046-minus-0000041", "task": "Serve meals", "image": [ "images/agibotworld/455/721837/cam_high/frame_000050.jpg", "images/agibotworld/455/721837/cam_high/frame_001548.jpg", "images/agibotworld/455/715725/cam_high/frame_000226.jpg", "images/agibotworld/455/715725/cam_left_wrist/frame_000226.jpg", "images/agibotworld/455/715725/cam_right_wrist/frame_000226.jpg", "images/agibotworld/455/715725/cam_high/frame_000046.jpg", "images/agibotworld/455/715725/cam_left_wrist/frame_000046.jpg", "images/agibotworld/455/715725/cam_right_wrist/frame_000046.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Serve meals`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Serve meals`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_695_863018-train-bf_000959-af_000839-minus-0000538", "task": "Turn on the fan", "image": [ "images/agibotworld/695/866075/cam_high/frame_000030.jpg", "images/agibotworld/695/866075/cam_high/frame_001183.jpg", "images/agibotworld/695/863018/cam_high/frame_000959.jpg", "images/agibotworld/695/863018/cam_left_wrist/frame_000959.jpg", "images/agibotworld/695/863018/cam_right_wrist/frame_000959.jpg", "images/agibotworld/695/863018/cam_high/frame_000839.jpg", "images/agibotworld/695/863018/cam_left_wrist/frame_000839.jpg", "images/agibotworld/695/863018/cam_right_wrist/frame_000839.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Turn on the fan`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Turn on the fan`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_389_691031-train-bf_000716-af_000581-minus-0000518", "task": "Pickup in the supermarket", "image": [ "images/agibotworld/389/691364/cam_high/frame_000000.jpg", "images/agibotworld/389/691364/cam_high/frame_000975.jpg", "images/agibotworld/389/691031/cam_high/frame_000716.jpg", "images/agibotworld/389/691031/cam_left_wrist/frame_000716.jpg", "images/agibotworld/389/691031/cam_right_wrist/frame_000716.jpg", "images/agibotworld/389/691031/cam_high/frame_000581.jpg", "images/agibotworld/389/691031/cam_left_wrist/frame_000581.jpg", "images/agibotworld/389/691031/cam_right_wrist/frame_000581.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pickup in the supermarket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pickup in the supermarket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -27.8, "hop_absolute_value": 27.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_492_724003-train-bf_001857-af_000120-minus-0000053", "task": "Pack in the supermarket cashier", "image": [ "images/agibotworld/492/724003/cam_high/frame_000000.jpg", "images/agibotworld/492/724003/cam_high/frame_003219.jpg", "images/agibotworld/492/724003/cam_high/frame_001857.jpg", "images/agibotworld/492/724003/cam_left_wrist/frame_001857.jpg", "images/agibotworld/492/724003/cam_right_wrist/frame_001857.jpg", "images/agibotworld/492/724003/cam_high/frame_000120.jpg", "images/agibotworld/492/724003/cam_left_wrist/frame_000120.jpg", "images/agibotworld/492/724003/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pack in the supermarket cashier`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pack in the supermarket cashier`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -90.6, "hop_absolute_value": 90.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_389_691314-train-bf_000388-af_000176-minus-0000126", "task": "Pickup in the supermarket", "image": [ "images/agibotworld/389/691069/cam_high/frame_000000.jpg", "images/agibotworld/389/691069/cam_high/frame_000902.jpg", "images/agibotworld/389/691314/cam_high/frame_000388.jpg", "images/agibotworld/389/691314/cam_left_wrist/frame_000388.jpg", "images/agibotworld/389/691314/cam_right_wrist/frame_000388.jpg", "images/agibotworld/389/691314/cam_high/frame_000176.jpg", "images/agibotworld/389/691314/cam_left_wrist/frame_000176.jpg", "images/agibotworld/389/691314/cam_right_wrist/frame_000176.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pickup in the supermarket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pickup in the supermarket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -57.1, "hop_absolute_value": 57.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_438_691345-train-bf_000853-af_000805-minus-0000258", "task": "Place the pen into the pen holder.", "image": [ "images/agibotworld/438/714290/cam_high/frame_000031.jpg", "images/agibotworld/438/714290/cam_high/frame_000885.jpg", "images/agibotworld/438/691345/cam_high/frame_000853.jpg", "images/agibotworld/438/691345/cam_left_wrist/frame_000853.jpg", "images/agibotworld/438/691345/cam_right_wrist/frame_000853.jpg", "images/agibotworld/438/691345/cam_high/frame_000805.jpg", "images/agibotworld/438/691345/cam_left_wrist/frame_000805.jpg", "images/agibotworld/438/691345/cam_right_wrist/frame_000805.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the pen into the pen holder.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the pen into the pen holder.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.7, "hop_absolute_value": 7.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_715_881604-train-bf_000831-af_000653-minus-0000167", "task": "Insert the key and open the door.", "image": [ "images/agibotworld/715/894318/cam_high/frame_000000.jpg", "images/agibotworld/715/894318/cam_high/frame_001298.jpg", "images/agibotworld/715/881604/cam_high/frame_000831.jpg", "images/agibotworld/715/881604/cam_left_wrist/frame_000831.jpg", "images/agibotworld/715/881604/cam_right_wrist/frame_000831.jpg", "images/agibotworld/715/881604/cam_high/frame_000653.jpg", "images/agibotworld/715/881604/cam_left_wrist/frame_000653.jpg", "images/agibotworld/715/881604/cam_right_wrist/frame_000653.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert the key and open the door.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert the key and open the door.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -41.7, "hop_absolute_value": 41.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_367_652746-train-bf_000776-af_000046-minus-0000013", "task": "Take toast from toaster", "image": [ "images/agibotworld/367/649878/cam_high/frame_000022.jpg", "images/agibotworld/367/649878/cam_high/frame_000823.jpg", "images/agibotworld/367/652746/cam_high/frame_000776.jpg", "images/agibotworld/367/652746/cam_left_wrist/frame_000776.jpg", "images/agibotworld/367/652746/cam_right_wrist/frame_000776.jpg", "images/agibotworld/367/652746/cam_high/frame_000046.jpg", "images/agibotworld/367/652746/cam_left_wrist/frame_000046.jpg", "images/agibotworld/367/652746/cam_right_wrist/frame_000046.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take toast from toaster`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take toast from toaster`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -94.7, "hop_absolute_value": 94.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_773_909434-train-bf_002140-af_001325-minus-0000259", "task": "Milk the cow", "image": [ "images/agibotworld/773/915664/cam_high/frame_000000.jpg", "images/agibotworld/773/915664/cam_high/frame_002288.jpg", "images/agibotworld/773/909434/cam_high/frame_002140.jpg", "images/agibotworld/773/909434/cam_left_wrist/frame_002140.jpg", "images/agibotworld/773/909434/cam_right_wrist/frame_002140.jpg", "images/agibotworld/773/909434/cam_high/frame_001325.jpg", "images/agibotworld/773/909434/cam_left_wrist/frame_001325.jpg", "images/agibotworld/773/909434/cam_right_wrist/frame_001325.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Milk the cow`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Milk the cow`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -32.9, "hop_absolute_value": 32.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_440_752908-train-bf_001892-af_001684-minus-0000808", "task": "Iron clothes", "image": [ "images/agibotworld/440/697279/cam_high/frame_000000.jpg", "images/agibotworld/440/697279/cam_high/frame_002649.jpg", "images/agibotworld/440/752908/cam_high/frame_001892.jpg", "images/agibotworld/440/752908/cam_left_wrist/frame_001892.jpg", "images/agibotworld/440/752908/cam_right_wrist/frame_001892.jpg", "images/agibotworld/440/752908/cam_high/frame_001684.jpg", "images/agibotworld/440/752908/cam_left_wrist/frame_001684.jpg", "images/agibotworld/440/752908/cam_right_wrist/frame_001684.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Iron clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Iron clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -19.6, "hop_absolute_value": 19.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_541_771911-train-bf_000418-af_000084-minus-0000036", "task": "Clean toilet with a rag", "image": [ "images/agibotworld/541/771911/cam_high/frame_000000.jpg", "images/agibotworld/541/771911/cam_high/frame_001451.jpg", "images/agibotworld/541/771911/cam_high/frame_000418.jpg", "images/agibotworld/541/771911/cam_left_wrist/frame_000418.jpg", "images/agibotworld/541/771911/cam_right_wrist/frame_000418.jpg", "images/agibotworld/541/771911/cam_high/frame_000084.jpg", "images/agibotworld/541/771911/cam_left_wrist/frame_000084.jpg", "images/agibotworld/541/771911/cam_right_wrist/frame_000084.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Clean toilet with a rag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Clean toilet with a rag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_597_826536-train-bf_001030-af_000597-minus-0000565", "task": "Restock the hanging basket area", "image": [ "images/agibotworld/597/826536/cam_high/frame_000000.jpg", "images/agibotworld/597/826536/cam_high/frame_001030.jpg", "images/agibotworld/597/826536/cam_high/frame_001030.jpg", "images/agibotworld/597/826536/cam_left_wrist/frame_001030.jpg", "images/agibotworld/597/826536/cam_right_wrist/frame_001030.jpg", "images/agibotworld/597/826536/cam_high/frame_000597.jpg", "images/agibotworld/597/826536/cam_left_wrist/frame_000597.jpg", "images/agibotworld/597/826536/cam_right_wrist/frame_000597.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Restock the hanging basket area`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Restock the hanging basket area`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -45.8, "hop_absolute_value": 45.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_478_700748-train-bf_000985-af_000615-minus-0000215", "task": "Wipe the mirror cabinet", "image": [ "images/agibotworld/478/698409/cam_high/frame_000000.jpg", "images/agibotworld/478/698409/cam_high/frame_002219.jpg", "images/agibotworld/478/700748/cam_high/frame_000985.jpg", "images/agibotworld/478/700748/cam_left_wrist/frame_000985.jpg", "images/agibotworld/478/700748/cam_right_wrist/frame_000985.jpg", "images/agibotworld/478/700748/cam_high/frame_000615.jpg", "images/agibotworld/478/700748/cam_left_wrist/frame_000615.jpg", "images/agibotworld/478/700748/cam_right_wrist/frame_000615.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wipe the mirror cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wipe the mirror cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -42.1, "hop_absolute_value": 42.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_508_774760-train-bf_001062-af_001004-minus-0000663", "task": "Place items in the bag.", "image": [ "images/agibotworld/508/762878/cam_high/frame_000000.jpg", "images/agibotworld/508/762878/cam_high/frame_001459.jpg", "images/agibotworld/508/774760/cam_high/frame_001062.jpg", "images/agibotworld/508/774760/cam_left_wrist/frame_001062.jpg", "images/agibotworld/508/774760/cam_right_wrist/frame_001062.jpg", "images/agibotworld/508/774760/cam_high/frame_001004.jpg", "images/agibotworld/508/774760/cam_left_wrist/frame_001004.jpg", "images/agibotworld/508/774760/cam_right_wrist/frame_001004.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place items in the bag.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place items in the bag.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.1, "hop_absolute_value": 7.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_462_856019-train-bf_001238-af_000676-minus-0000196", "task": "Discard the trash bag into the large bin.", "image": [ "images/agibotworld/462/870579/cam_high/frame_000000.jpg", "images/agibotworld/462/870579/cam_high/frame_001476.jpg", "images/agibotworld/462/856019/cam_high/frame_001238.jpg", "images/agibotworld/462/856019/cam_left_wrist/frame_001238.jpg", "images/agibotworld/462/856019/cam_right_wrist/frame_001238.jpg", "images/agibotworld/462/856019/cam_high/frame_000676.jpg", "images/agibotworld/462/856019/cam_left_wrist/frame_000676.jpg", "images/agibotworld/462/856019/cam_right_wrist/frame_000676.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Discard the trash bag into the large bin.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Discard the trash bag into the large bin.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -48.3, "hop_absolute_value": 48.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_541_774235-train-bf_001008-af_000685-minus-0000325", "task": "Clean toilet with a rag", "image": [ "images/agibotworld/541/771705/cam_high/frame_000000.jpg", "images/agibotworld/541/771705/cam_high/frame_001468.jpg", "images/agibotworld/541/774235/cam_high/frame_001008.jpg", "images/agibotworld/541/774235/cam_left_wrist/frame_001008.jpg", "images/agibotworld/541/774235/cam_right_wrist/frame_001008.jpg", "images/agibotworld/541/774235/cam_high/frame_000685.jpg", "images/agibotworld/541/774235/cam_left_wrist/frame_000685.jpg", "images/agibotworld/541/774235/cam_right_wrist/frame_000685.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Clean toilet with a rag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Clean toilet with a rag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -18.8, "hop_absolute_value": 18.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_357_654077-train-bf_001487-af_000281-minus-0000043", "task": "Wash dishes with dishwasher", "image": [ "images/agibotworld/357/662433/cam_high/frame_000000.jpg", "images/agibotworld/357/662433/cam_high/frame_003023.jpg", "images/agibotworld/357/654077/cam_high/frame_001487.jpg", "images/agibotworld/357/654077/cam_left_wrist/frame_001487.jpg", "images/agibotworld/357/654077/cam_right_wrist/frame_001487.jpg", "images/agibotworld/357/654077/cam_high/frame_000281.jpg", "images/agibotworld/357/654077/cam_left_wrist/frame_000281.jpg", "images/agibotworld/357/654077/cam_right_wrist/frame_000281.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wash dishes with dishwasher`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wash dishes with dishwasher`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -92.0, "hop_absolute_value": 92.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_365_692455-train-bf_000643-af_000239-minus-0000851", "task": "Sort personal care products", "image": [ "images/agibotworld/365/692455/cam_high/frame_000000.jpg", "images/agibotworld/365/692455/cam_high/frame_001156.jpg", "images/agibotworld/365/692455/cam_high/frame_000643.jpg", "images/agibotworld/365/692455/cam_left_wrist/frame_000643.jpg", "images/agibotworld/365/692455/cam_right_wrist/frame_000643.jpg", "images/agibotworld/365/692455/cam_high/frame_000239.jpg", "images/agibotworld/365/692455/cam_left_wrist/frame_000239.jpg", "images/agibotworld/365/692455/cam_right_wrist/frame_000239.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort personal care products`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort personal care products`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_570_807204-train-bf_001464-af_000607-minus-0000074", "task": "Fold the T-shirt on the field", "image": [ "images/agibotworld/570/812055/cam_high/frame_000000.jpg", "images/agibotworld/570/812055/cam_high/frame_002203.jpg", "images/agibotworld/570/807204/cam_high/frame_001464.jpg", "images/agibotworld/570/807204/cam_left_wrist/frame_001464.jpg", "images/agibotworld/570/807204/cam_right_wrist/frame_001464.jpg", "images/agibotworld/570/807204/cam_high/frame_000607.jpg", "images/agibotworld/570/807204/cam_left_wrist/frame_000607.jpg", "images/agibotworld/570/807204/cam_right_wrist/frame_000607.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the T-shirt on the field`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the T-shirt on the field`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_574_818754-train-bf_001351-af_001265-minus-0000273", "task": "Produce ice with an ice maker", "image": [ "images/agibotworld/574/808287/cam_high/frame_000000.jpg", "images/agibotworld/574/808287/cam_high/frame_001227.jpg", "images/agibotworld/574/818754/cam_high/frame_001351.jpg", "images/agibotworld/574/818754/cam_left_wrist/frame_001351.jpg", "images/agibotworld/574/818754/cam_right_wrist/frame_001351.jpg", "images/agibotworld/574/818754/cam_high/frame_001265.jpg", "images/agibotworld/574/818754/cam_left_wrist/frame_001265.jpg", "images/agibotworld/574/818754/cam_right_wrist/frame_001265.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Produce ice with an ice maker`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Produce ice with an ice maker`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -8.0, "hop_absolute_value": 8.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_428_771348-train-bf_000746-af_000000-minus-0000007", "task": "Open drawer and store items", "image": [ "images/agibotworld/428/771348/cam_high/frame_000000.jpg", "images/agibotworld/428/771348/cam_high/frame_002078.jpg", "images/agibotworld/428/771348/cam_high/frame_000746.jpg", "images/agibotworld/428/771348/cam_left_wrist/frame_000746.jpg", "images/agibotworld/428/771348/cam_right_wrist/frame_000746.jpg", "images/agibotworld/428/771348/cam_high/frame_000000.jpg", "images/agibotworld/428/771348/cam_left_wrist/frame_000000.jpg", "images/agibotworld/428/771348/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open drawer and store items`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open drawer and store items`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_561_821983-train-bf_000959-af_000720-minus-0000122", "task": "Flatten and fold shorts", "image": [ "images/agibotworld/561/833063/cam_high/frame_000000.jpg", "images/agibotworld/561/833063/cam_high/frame_000964.jpg", "images/agibotworld/561/821983/cam_high/frame_000959.jpg", "images/agibotworld/561/821983/cam_left_wrist/frame_000959.jpg", "images/agibotworld/561/821983/cam_right_wrist/frame_000959.jpg", "images/agibotworld/561/821983/cam_high/frame_000720.jpg", "images/agibotworld/561/821983/cam_left_wrist/frame_000720.jpg", "images/agibotworld/561/821983/cam_right_wrist/frame_000720.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Flatten and fold shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Flatten and fold shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -35.7, "hop_absolute_value": 35.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_483_751114-train-bf_002241-af_000052-minus-0000019", "task": "Place items from the meeting room table into the storage box.", "image": [ "images/agibotworld/483/751114/cam_high/frame_000000.jpg", "images/agibotworld/483/751114/cam_high/frame_003375.jpg", "images/agibotworld/483/751114/cam_high/frame_002241.jpg", "images/agibotworld/483/751114/cam_left_wrist/frame_002241.jpg", "images/agibotworld/483/751114/cam_right_wrist/frame_002241.jpg", "images/agibotworld/483/751114/cam_high/frame_000052.jpg", "images/agibotworld/483/751114/cam_left_wrist/frame_000052.jpg", "images/agibotworld/483/751114/cam_right_wrist/frame_000052.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place items from the meeting room table into the storage box.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place items from the meeting room table into the storage box.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -97.7, "hop_absolute_value": 97.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_466_830784-train-bf_001661-af_000355-minus-0000109", "task": "Separate dark and light clothes", "image": [ "images/agibotworld/466/809957/cam_high/frame_000000.jpg", "images/agibotworld/466/809957/cam_high/frame_002270.jpg", "images/agibotworld/466/830784/cam_high/frame_001661.jpg", "images/agibotworld/466/830784/cam_left_wrist/frame_001661.jpg", "images/agibotworld/466/830784/cam_right_wrist/frame_001661.jpg", "images/agibotworld/466/830784/cam_high/frame_000355.jpg", "images/agibotworld/466/830784/cam_left_wrist/frame_000355.jpg", "images/agibotworld/466/830784/cam_right_wrist/frame_000355.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Separate dark and light clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Separate dark and light clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_748_893216-train-bf_000679-af_000611-minus-0000745", "task": "Scan for security check", "image": [ "images/agibotworld/748/893272/cam_high/frame_000000.jpg", "images/agibotworld/748/893272/cam_high/frame_002812.jpg", "images/agibotworld/748/893216/cam_high/frame_000679.jpg", "images/agibotworld/748/893216/cam_left_wrist/frame_000679.jpg", "images/agibotworld/748/893216/cam_right_wrist/frame_000679.jpg", "images/agibotworld/748/893216/cam_high/frame_000611.jpg", "images/agibotworld/748/893216/cam_left_wrist/frame_000611.jpg", "images/agibotworld/748/893216/cam_right_wrist/frame_000611.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Scan for security check`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Scan for security check`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_360_675924-train-bf_001023-af_000878-minus-0000646", "task": "Packing in the supermarket", "image": [ "images/agibotworld/360/674981/cam_high/frame_000012.jpg", "images/agibotworld/360/674981/cam_high/frame_001037.jpg", "images/agibotworld/360/675924/cam_high/frame_001023.jpg", "images/agibotworld/360/675924/cam_left_wrist/frame_001023.jpg", "images/agibotworld/360/675924/cam_right_wrist/frame_001023.jpg", "images/agibotworld/360/675924/cam_high/frame_000878.jpg", "images/agibotworld/360/675924/cam_left_wrist/frame_000878.jpg", "images/agibotworld/360/675924/cam_right_wrist/frame_000878.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Packing in the supermarket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Packing in the supermarket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_732_891682-train-bf_001557-af_000371-minus-0000071", "task": "Roll the dough", "image": [ "images/agibotworld/732/910511/cam_high/frame_000000.jpg", "images/agibotworld/732/910511/cam_high/frame_001180.jpg", "images/agibotworld/732/891682/cam_high/frame_001557.jpg", "images/agibotworld/732/891682/cam_left_wrist/frame_001557.jpg", "images/agibotworld/732/891682/cam_right_wrist/frame_001557.jpg", "images/agibotworld/732/891682/cam_high/frame_000371.jpg", "images/agibotworld/732/891682/cam_left_wrist/frame_000371.jpg", "images/agibotworld/732/891682/cam_right_wrist/frame_000371.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Roll the dough`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Roll the dough`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -81.8, "hop_absolute_value": 81.8, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_544_819003-train-bf_000821-af_000292-minus-0000119", "task": "Wash the pot and spatula", "image": [ "images/agibotworld/544/821409/cam_high/frame_000000.jpg", "images/agibotworld/544/821409/cam_high/frame_002860.jpg", "images/agibotworld/544/819003/cam_high/frame_000821.jpg", "images/agibotworld/544/819003/cam_left_wrist/frame_000821.jpg", "images/agibotworld/544/819003/cam_right_wrist/frame_000821.jpg", "images/agibotworld/544/819003/cam_high/frame_000292.jpg", "images/agibotworld/544/819003/cam_left_wrist/frame_000292.jpg", "images/agibotworld/544/819003/cam_right_wrist/frame_000292.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wash the pot and spatula`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wash the pot and spatula`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.9, "hop_absolute_value": 73.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_507_739368-train-bf_003122-af_002477-minus-0000273", "task": "Prepare breakfast", "image": [ "images/agibotworld/507/739368/cam_high/frame_000041.jpg", "images/agibotworld/507/739368/cam_high/frame_003192.jpg", "images/agibotworld/507/739368/cam_high/frame_003122.jpg", "images/agibotworld/507/739368/cam_left_wrist/frame_003122.jpg", "images/agibotworld/507/739368/cam_right_wrist/frame_003122.jpg", "images/agibotworld/507/739368/cam_high/frame_002477.jpg", "images/agibotworld/507/739368/cam_left_wrist/frame_002477.jpg", "images/agibotworld/507/739368/cam_right_wrist/frame_002477.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Prepare breakfast`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Prepare breakfast`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -26.8, "hop_absolute_value": 26.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_584_813005-train-bf_001107-af_000630-minus-0000225", "task": "Pack permanent magnet ingots", "image": [ "images/agibotworld/584/812928/cam_high/frame_000000.jpg", "images/agibotworld/584/812928/cam_high/frame_001192.jpg", "images/agibotworld/584/813005/cam_high/frame_001107.jpg", "images/agibotworld/584/813005/cam_left_wrist/frame_001107.jpg", "images/agibotworld/584/813005/cam_right_wrist/frame_001107.jpg", "images/agibotworld/584/813005/cam_high/frame_000630.jpg", "images/agibotworld/584/813005/cam_left_wrist/frame_000630.jpg", "images/agibotworld/584/813005/cam_right_wrist/frame_000630.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pack permanent magnet ingots`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pack permanent magnet ingots`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -36.7, "hop_absolute_value": 36.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_367_652673-train-bf_000372-af_000302-minus-0000517", "task": "Take toast from toaster", "image": [ "images/agibotworld/367/649454/cam_high/frame_000035.jpg", "images/agibotworld/367/649454/cam_high/frame_000717.jpg", "images/agibotworld/367/652673/cam_high/frame_000372.jpg", "images/agibotworld/367/652673/cam_left_wrist/frame_000372.jpg", "images/agibotworld/367/652673/cam_right_wrist/frame_000372.jpg", "images/agibotworld/367/652673/cam_high/frame_000302.jpg", "images/agibotworld/367/652673/cam_left_wrist/frame_000302.jpg", "images/agibotworld/367/652673/cam_right_wrist/frame_000302.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take toast from toaster`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take toast from toaster`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -22.2, "hop_absolute_value": 22.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_683_871277-train-bf_002298-af_002223-minus-0000280", "task": "Serve the meal", "image": [ "images/agibotworld/683/868698/cam_high/frame_000000.jpg", "images/agibotworld/683/868698/cam_high/frame_001879.jpg", "images/agibotworld/683/871277/cam_high/frame_002298.jpg", "images/agibotworld/683/871277/cam_left_wrist/frame_002298.jpg", "images/agibotworld/683/871277/cam_right_wrist/frame_002298.jpg", "images/agibotworld/683/871277/cam_high/frame_002223.jpg", "images/agibotworld/683/871277/cam_left_wrist/frame_002223.jpg", "images/agibotworld/683/871277/cam_right_wrist/frame_002223.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Serve the meal`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Serve the meal`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.0, "hop_absolute_value": 3.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_521_768875-train-bf_000984-af_000671-minus-0000917", "task": "Dispose of the trash on the desk.", "image": [ "images/agibotworld/521/758125/cam_high/frame_000032.jpg", "images/agibotworld/521/758125/cam_high/frame_001225.jpg", "images/agibotworld/521/768875/cam_high/frame_000984.jpg", "images/agibotworld/521/768875/cam_left_wrist/frame_000984.jpg", "images/agibotworld/521/768875/cam_right_wrist/frame_000984.jpg", "images/agibotworld/521/768875/cam_high/frame_000671.jpg", "images/agibotworld/521/768875/cam_left_wrist/frame_000671.jpg", "images/agibotworld/521/768875/cam_right_wrist/frame_000671.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of the trash on the desk.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of the trash on the desk.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -41.4, "hop_absolute_value": 41.4, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_785_925850-train-bf_003748-af_003694-minus-0000296", "task": "Serve the meal", "image": [ "images/agibotworld/785/919993/cam_high/frame_000065.jpg", "images/agibotworld/785/919993/cam_high/frame_004321.jpg", "images/agibotworld/785/925850/cam_high/frame_003748.jpg", "images/agibotworld/785/925850/cam_left_wrist/frame_003748.jpg", "images/agibotworld/785/925850/cam_right_wrist/frame_003748.jpg", "images/agibotworld/785/925850/cam_high/frame_003694.jpg", "images/agibotworld/785/925850/cam_left_wrist/frame_003694.jpg", "images/agibotworld/785/925850/cam_right_wrist/frame_003694.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Serve the meal`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Serve the meal`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.6, "hop_absolute_value": 1.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_722_889273-train-bf_002138-af_001910-minus-0000689", "task": "Tighten the bottle cap", "image": [ "images/agibotworld/722/886117/cam_high/frame_000000.jpg", "images/agibotworld/722/886117/cam_high/frame_002661.jpg", "images/agibotworld/722/889273/cam_high/frame_002138.jpg", "images/agibotworld/722/889273/cam_left_wrist/frame_002138.jpg", "images/agibotworld/722/889273/cam_right_wrist/frame_002138.jpg", "images/agibotworld/722/889273/cam_high/frame_001910.jpg", "images/agibotworld/722/889273/cam_left_wrist/frame_001910.jpg", "images/agibotworld/722/889273/cam_right_wrist/frame_001910.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Tighten the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Tighten the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_376_650750-train-bf_000501-af_000473-minus-0000546", "task": "Sort electronic products", "image": [ "images/agibotworld/376/680052/cam_high/frame_000000.jpg", "images/agibotworld/376/680052/cam_high/frame_000912.jpg", "images/agibotworld/376/650750/cam_high/frame_000501.jpg", "images/agibotworld/376/650750/cam_left_wrist/frame_000501.jpg", "images/agibotworld/376/650750/cam_right_wrist/frame_000501.jpg", "images/agibotworld/376/650750/cam_high/frame_000473.jpg", "images/agibotworld/376/650750/cam_left_wrist/frame_000473.jpg", "images/agibotworld/376/650750/cam_right_wrist/frame_000473.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort electronic products`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort electronic products`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.1, "hop_absolute_value": 7.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_483_723441-train-bf_002159-af_001975-minus-0000824", "task": "Place items from the meeting room table into the storage box.", "image": [ "images/agibotworld/483/748921/cam_high/frame_000000.jpg", "images/agibotworld/483/748921/cam_high/frame_003127.jpg", "images/agibotworld/483/723441/cam_high/frame_002159.jpg", "images/agibotworld/483/723441/cam_left_wrist/frame_002159.jpg", "images/agibotworld/483/723441/cam_right_wrist/frame_002159.jpg", "images/agibotworld/483/723441/cam_high/frame_001975.jpg", "images/agibotworld/483/723441/cam_left_wrist/frame_001975.jpg", "images/agibotworld/483/723441/cam_right_wrist/frame_001975.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place items from the meeting room table into the storage box.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place items from the meeting room table into the storage box.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -8.9, "hop_absolute_value": 8.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_544_804897-train-bf_002120-af_002011-minus-0000688", "task": "Wash the pot and spatula", "image": [ "images/agibotworld/544/818755/cam_high/frame_000000.jpg", "images/agibotworld/544/818755/cam_high/frame_003505.jpg", "images/agibotworld/544/804897/cam_high/frame_002120.jpg", "images/agibotworld/544/804897/cam_left_wrist/frame_002120.jpg", "images/agibotworld/544/804897/cam_right_wrist/frame_002120.jpg", "images/agibotworld/544/804897/cam_high/frame_002011.jpg", "images/agibotworld/544/804897/cam_left_wrist/frame_002011.jpg", "images/agibotworld/544/804897/cam_right_wrist/frame_002011.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wash the pot and spatula`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wash the pot and spatula`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.0, "hop_absolute_value": 2.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_485_715970-train-bf_001820-af_000947-minus-0000777", "task": "Pick up the scrap paper on the desk and feed it into the shredder for destruction.", "image": [ "images/agibotworld/485/917998/cam_high/frame_000000.jpg", "images/agibotworld/485/917998/cam_high/frame_001220.jpg", "images/agibotworld/485/715970/cam_high/frame_001820.jpg", "images/agibotworld/485/715970/cam_left_wrist/frame_001820.jpg", "images/agibotworld/485/715970/cam_right_wrist/frame_001820.jpg", "images/agibotworld/485/715970/cam_high/frame_000947.jpg", "images/agibotworld/485/715970/cam_left_wrist/frame_000947.jpg", "images/agibotworld/485/715970/cam_right_wrist/frame_000947.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the scrap paper on the desk and feed it into the shredder for destruction.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the scrap paper on the desk and feed it into the shredder for destruction.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -44.8, "hop_absolute_value": 44.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_527_784473-train-bf_004638-af_001263-minus-0000139", "task": "Arrange sofa", "image": [ "images/agibotworld/527/756146/cam_high/frame_000000.jpg", "images/agibotworld/527/756146/cam_high/frame_004657.jpg", "images/agibotworld/527/784473/cam_high/frame_004638.jpg", "images/agibotworld/527/784473/cam_left_wrist/frame_004638.jpg", "images/agibotworld/527/784473/cam_right_wrist/frame_004638.jpg", "images/agibotworld/527/784473/cam_high/frame_001263.jpg", "images/agibotworld/527/784473/cam_left_wrist/frame_001263.jpg", "images/agibotworld/527/784473/cam_right_wrist/frame_001263.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Arrange sofa`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Arrange sofa`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -71.2, "hop_absolute_value": 71.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_613_841922-train-bf_000397-af_000141-minus-0000128", "task": "Insert the pen cap", "image": [ "images/agibotworld/613/843301/cam_high/frame_000000.jpg", "images/agibotworld/613/843301/cam_high/frame_001354.jpg", "images/agibotworld/613/841922/cam_high/frame_000397.jpg", "images/agibotworld/613/841922/cam_left_wrist/frame_000397.jpg", "images/agibotworld/613/841922/cam_right_wrist/frame_000397.jpg", "images/agibotworld/613/841922/cam_high/frame_000141.jpg", "images/agibotworld/613/841922/cam_left_wrist/frame_000141.jpg", "images/agibotworld/613/841922/cam_right_wrist/frame_000141.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert the pen cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert the pen cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_486_749916-train-bf_002158-af_002080-minus-0000740", "task": "Remove bottled water from the carton and arrange it neatly on the table.", "image": [ "images/agibotworld/486/755581/cam_high/frame_000000.jpg", "images/agibotworld/486/755581/cam_high/frame_003317.jpg", "images/agibotworld/486/749916/cam_high/frame_002158.jpg", "images/agibotworld/486/749916/cam_left_wrist/frame_002158.jpg", "images/agibotworld/486/749916/cam_right_wrist/frame_002158.jpg", "images/agibotworld/486/749916/cam_high/frame_002080.jpg", "images/agibotworld/486/749916/cam_left_wrist/frame_002080.jpg", "images/agibotworld/486/749916/cam_right_wrist/frame_002080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove bottled water from the carton and arrange it neatly on the table.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove bottled water from the carton and arrange it neatly on the table.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.1, "hop_absolute_value": 4.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_773_893146-train-bf_000503-af_000071-minus-0000040", "task": "Milk the cow", "image": [ "images/agibotworld/773/906709/cam_high/frame_000000.jpg", "images/agibotworld/773/906709/cam_high/frame_002460.jpg", "images/agibotworld/773/893146/cam_high/frame_000503.jpg", "images/agibotworld/773/893146/cam_left_wrist/frame_000503.jpg", "images/agibotworld/773/893146/cam_right_wrist/frame_000503.jpg", "images/agibotworld/773/893146/cam_high/frame_000071.jpg", "images/agibotworld/773/893146/cam_left_wrist/frame_000071.jpg", "images/agibotworld/773/893146/cam_right_wrist/frame_000071.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Milk the cow`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Milk the cow`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -90.5, "hop_absolute_value": 90.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_732_883627-train-bf_000841-af_000510-minus-0000154", "task": "Roll the dough", "image": [ "images/agibotworld/732/900231/cam_high/frame_000029.jpg", "images/agibotworld/732/900231/cam_high/frame_001727.jpg", "images/agibotworld/732/883627/cam_high/frame_000841.jpg", "images/agibotworld/732/883627/cam_left_wrist/frame_000841.jpg", "images/agibotworld/732/883627/cam_right_wrist/frame_000841.jpg", "images/agibotworld/732/883627/cam_high/frame_000510.jpg", "images/agibotworld/732/883627/cam_left_wrist/frame_000510.jpg", "images/agibotworld/732/883627/cam_right_wrist/frame_000510.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Roll the dough`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Roll the dough`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -47.1, "hop_absolute_value": 47.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_563_901663-train-bf_002784-af_000988-minus-0000112", "task": "Remove the baked dessert from the oven.", "image": [ "images/agibotworld/563/840968/cam_high/frame_000035.jpg", "images/agibotworld/563/840968/cam_high/frame_002699.jpg", "images/agibotworld/563/901663/cam_high/frame_002784.jpg", "images/agibotworld/563/901663/cam_left_wrist/frame_002784.jpg", "images/agibotworld/563/901663/cam_right_wrist/frame_002784.jpg", "images/agibotworld/563/901663/cam_high/frame_000988.jpg", "images/agibotworld/563/901663/cam_left_wrist/frame_000988.jpg", "images/agibotworld/563/901663/cam_right_wrist/frame_000988.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the baked dessert from the oven.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the baked dessert from the oven.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -80.6, "hop_absolute_value": 80.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_596_822125-train-bf_000913-af_000704-minus-0000678", "task": "Place the inner pot with rice into the rice cooker to cook.", "image": [ "images/agibotworld/596/821839/cam_high/frame_000035.jpg", "images/agibotworld/596/821839/cam_high/frame_001759.jpg", "images/agibotworld/596/822125/cam_high/frame_000913.jpg", "images/agibotworld/596/822125/cam_left_wrist/frame_000913.jpg", "images/agibotworld/596/822125/cam_right_wrist/frame_000913.jpg", "images/agibotworld/596/822125/cam_high/frame_000704.jpg", "images/agibotworld/596/822125/cam_left_wrist/frame_000704.jpg", "images/agibotworld/596/822125/cam_right_wrist/frame_000704.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the inner pot with rice into the rice cooker to cook.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the inner pot with rice into the rice cooker to cook.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_575_806206-train-bf_000564-af_000133-minus-0000075", "task": "Store toys", "image": [ "images/agibotworld/575/807540/cam_high/frame_000000.jpg", "images/agibotworld/575/807540/cam_high/frame_000526.jpg", "images/agibotworld/575/806206/cam_high/frame_000564.jpg", "images/agibotworld/575/806206/cam_left_wrist/frame_000564.jpg", "images/agibotworld/575/806206/cam_right_wrist/frame_000564.jpg", "images/agibotworld/575/806206/cam_high/frame_000133.jpg", "images/agibotworld/575/806206/cam_left_wrist/frame_000133.jpg", "images/agibotworld/575/806206/cam_right_wrist/frame_000133.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Store toys`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Store toys`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -77.8, "hop_absolute_value": 77.8, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_351_832505-train-bf_000682-af_000534-minus-0000749", "task": "Open the wardrobe and hang the clothes.", "image": [ "images/agibotworld/351/850860/cam_high/frame_000000.jpg", "images/agibotworld/351/850860/cam_high/frame_001304.jpg", "images/agibotworld/351/832505/cam_high/frame_000682.jpg", "images/agibotworld/351/832505/cam_left_wrist/frame_000682.jpg", "images/agibotworld/351/832505/cam_right_wrist/frame_000682.jpg", "images/agibotworld/351/832505/cam_high/frame_000534.jpg", "images/agibotworld/351/832505/cam_left_wrist/frame_000534.jpg", "images/agibotworld/351/832505/cam_right_wrist/frame_000534.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the wardrobe and hang the clothes.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the wardrobe and hang the clothes.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_712_868667-train-bf_000255-af_000074-minus-0000079", "task": "Carry bottled water.", "image": [ "images/agibotworld/712/877362/cam_high/frame_000000.jpg", "images/agibotworld/712/877362/cam_high/frame_001771.jpg", "images/agibotworld/712/868667/cam_high/frame_000255.jpg", "images/agibotworld/712/868667/cam_left_wrist/frame_000255.jpg", "images/agibotworld/712/868667/cam_right_wrist/frame_000255.jpg", "images/agibotworld/712/868667/cam_high/frame_000074.jpg", "images/agibotworld/712/868667/cam_left_wrist/frame_000074.jpg", "images/agibotworld/712/868667/cam_right_wrist/frame_000074.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Carry bottled water.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Carry bottled water.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -77.8, "hop_absolute_value": 77.8, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_454_701206-train-bf_001601-af_001566-minus-0000665", "task": "Make a sandwich", "image": [ "images/agibotworld/454/699907/cam_high/frame_000000.jpg", "images/agibotworld/454/699907/cam_high/frame_002089.jpg", "images/agibotworld/454/701206/cam_high/frame_001601.jpg", "images/agibotworld/454/701206/cam_left_wrist/frame_001601.jpg", "images/agibotworld/454/701206/cam_right_wrist/frame_001601.jpg", "images/agibotworld/454/701206/cam_high/frame_001566.jpg", "images/agibotworld/454/701206/cam_left_wrist/frame_001566.jpg", "images/agibotworld/454/701206/cam_right_wrist/frame_001566.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make a sandwich`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make a sandwich`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.9, "hop_absolute_value": 3.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_410_656325-train-bf_000348-af_000284-minus-0000142", "task": "Water Pouring in Restaurant", "image": [ "images/agibotworld/410/679315/cam_high/frame_000000.jpg", "images/agibotworld/410/679315/cam_high/frame_001119.jpg", "images/agibotworld/410/656325/cam_high/frame_000348.jpg", "images/agibotworld/410/656325/cam_left_wrist/frame_000348.jpg", "images/agibotworld/410/656325/cam_right_wrist/frame_000348.jpg", "images/agibotworld/410/656325/cam_high/frame_000284.jpg", "images/agibotworld/410/656325/cam_left_wrist/frame_000284.jpg", "images/agibotworld/410/656325/cam_right_wrist/frame_000284.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Water Pouring in Restaurant`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Water Pouring in Restaurant`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_376_677552-train-bf_000862-af_000268-minus-0000082", "task": "Sort electronic products", "image": [ "images/agibotworld/376/674023/cam_high/frame_000000.jpg", "images/agibotworld/376/674023/cam_high/frame_001240.jpg", "images/agibotworld/376/677552/cam_high/frame_000862.jpg", "images/agibotworld/376/677552/cam_left_wrist/frame_000862.jpg", "images/agibotworld/376/677552/cam_right_wrist/frame_000862.jpg", "images/agibotworld/376/677552/cam_high/frame_000268.jpg", "images/agibotworld/376/677552/cam_left_wrist/frame_000268.jpg", "images/agibotworld/376/677552/cam_right_wrist/frame_000268.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort electronic products`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort electronic products`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -76.2, "hop_absolute_value": 76.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_773_903268-train-bf_001900-af_001815-minus-0000773", "task": "Milk the cow", "image": [ "images/agibotworld/773/912373/cam_high/frame_000000.jpg", "images/agibotworld/773/912373/cam_high/frame_001709.jpg", "images/agibotworld/773/903268/cam_high/frame_001900.jpg", "images/agibotworld/773/903268/cam_left_wrist/frame_001900.jpg", "images/agibotworld/773/903268/cam_right_wrist/frame_001900.jpg", "images/agibotworld/773/903268/cam_high/frame_001815.jpg", "images/agibotworld/773/903268/cam_left_wrist/frame_001815.jpg", "images/agibotworld/773/903268/cam_right_wrist/frame_001815.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Milk the cow`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Milk the cow`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.1, "hop_absolute_value": 5.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_521_764193-train-bf_000684-af_000588-minus-0000750", "task": "Dispose of the trash on the desk.", "image": [ "images/agibotworld/521/757420/cam_high/frame_000000.jpg", "images/agibotworld/521/757420/cam_high/frame_001992.jpg", "images/agibotworld/521/764193/cam_high/frame_000684.jpg", "images/agibotworld/521/764193/cam_left_wrist/frame_000684.jpg", "images/agibotworld/521/764193/cam_right_wrist/frame_000684.jpg", "images/agibotworld/521/764193/cam_high/frame_000588.jpg", "images/agibotworld/521/764193/cam_left_wrist/frame_000588.jpg", "images/agibotworld/521/764193/cam_right_wrist/frame_000588.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of the trash on the desk.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of the trash on the desk.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -22.7, "hop_absolute_value": 22.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_480_721662-train-bf_000771-af_000384-minus-0000173", "task": "Make sandwiches with salad dressing 1213 edition", "image": [ "images/agibotworld/480/719096/cam_high/frame_000122.jpg", "images/agibotworld/480/719096/cam_high/frame_002519.jpg", "images/agibotworld/480/721662/cam_high/frame_000771.jpg", "images/agibotworld/480/721662/cam_left_wrist/frame_000771.jpg", "images/agibotworld/480/721662/cam_right_wrist/frame_000771.jpg", "images/agibotworld/480/721662/cam_high/frame_000384.jpg", "images/agibotworld/480/721662/cam_left_wrist/frame_000384.jpg", "images/agibotworld/480/721662/cam_right_wrist/frame_000384.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make sandwiches with salad dressing 1213 edition`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make sandwiches with salad dressing 1213 edition`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -54.5, "hop_absolute_value": 54.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_434_830284-train-bf_001141-af_000742-minus-0000228", "task": "Remove capsules from the coffee machine.", "image": [ "images/agibotworld/434/811045/cam_high/frame_000000.jpg", "images/agibotworld/434/811045/cam_high/frame_001798.jpg", "images/agibotworld/434/830284/cam_high/frame_001141.jpg", "images/agibotworld/434/830284/cam_left_wrist/frame_001141.jpg", "images/agibotworld/434/830284/cam_right_wrist/frame_001141.jpg", "images/agibotworld/434/830284/cam_high/frame_000742.jpg", "images/agibotworld/434/830284/cam_left_wrist/frame_000742.jpg", "images/agibotworld/434/830284/cam_right_wrist/frame_000742.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove capsules from the coffee machine.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove capsules from the coffee machine.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -36.0, "hop_absolute_value": 36.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_529_761150-train-bf_000654-af_000636-minus-0000495", "task": "Wipe the toilet with a cloth.", "image": [ "images/agibotworld/529/761329/cam_high/frame_000000.jpg", "images/agibotworld/529/761329/cam_high/frame_001764.jpg", "images/agibotworld/529/761150/cam_high/frame_000654.jpg", "images/agibotworld/529/761150/cam_left_wrist/frame_000654.jpg", "images/agibotworld/529/761150/cam_right_wrist/frame_000654.jpg", "images/agibotworld/529/761150/cam_high/frame_000636.jpg", "images/agibotworld/529/761150/cam_left_wrist/frame_000636.jpg", "images/agibotworld/529/761150/cam_right_wrist/frame_000636.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wipe the toilet with a cloth.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wipe the toilet with a cloth.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -18.2, "hop_absolute_value": 18.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_563_897082-train-bf_002248-af_000087-minus-0000018", "task": "Remove the baked dessert from the oven.", "image": [ "images/agibotworld/563/839242/cam_high/frame_000030.jpg", "images/agibotworld/563/839242/cam_high/frame_002474.jpg", "images/agibotworld/563/897082/cam_high/frame_002248.jpg", "images/agibotworld/563/897082/cam_left_wrist/frame_002248.jpg", "images/agibotworld/563/897082/cam_right_wrist/frame_002248.jpg", "images/agibotworld/563/897082/cam_high/frame_000087.jpg", "images/agibotworld/563/897082/cam_left_wrist/frame_000087.jpg", "images/agibotworld/563/897082/cam_right_wrist/frame_000087.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the baked dessert from the oven.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the baked dessert from the oven.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -98.0, "hop_absolute_value": 98.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_434_824344-train-bf_001204-af_000818-minus-0000653", "task": "Remove capsules from the coffee machine.", "image": [ "images/agibotworld/434/830816/cam_high/frame_000000.jpg", "images/agibotworld/434/830816/cam_high/frame_001379.jpg", "images/agibotworld/434/824344/cam_high/frame_001204.jpg", "images/agibotworld/434/824344/cam_left_wrist/frame_001204.jpg", "images/agibotworld/434/824344/cam_right_wrist/frame_001204.jpg", "images/agibotworld/434/824344/cam_high/frame_000818.jpg", "images/agibotworld/434/824344/cam_left_wrist/frame_000818.jpg", "images/agibotworld/434/824344/cam_right_wrist/frame_000818.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove capsules from the coffee machine.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove capsules from the coffee machine.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -23.3, "hop_absolute_value": 23.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_351_832354-train-bf_001297-af_000864-minus-0000908", "task": "Open the wardrobe and hang the clothes.", "image": [ "images/agibotworld/351/823743/cam_high/frame_000000.jpg", "images/agibotworld/351/823743/cam_high/frame_002276.jpg", "images/agibotworld/351/832354/cam_high/frame_001297.jpg", "images/agibotworld/351/832354/cam_left_wrist/frame_001297.jpg", "images/agibotworld/351/832354/cam_right_wrist/frame_001297.jpg", "images/agibotworld/351/832354/cam_high/frame_000864.jpg", "images/agibotworld/351/832354/cam_left_wrist/frame_000864.jpg", "images/agibotworld/351/832354/cam_right_wrist/frame_000864.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the wardrobe and hang the clothes.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the wardrobe and hang the clothes.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.9, "hop_absolute_value": 25.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_422_713721-train-bf_001375-af_000964-minus-0000259", "task": "Pack items for industrial logistics", "image": [ "images/agibotworld/422/701373/cam_high/frame_000000.jpg", "images/agibotworld/422/701373/cam_high/frame_002201.jpg", "images/agibotworld/422/713721/cam_high/frame_001375.jpg", "images/agibotworld/422/713721/cam_left_wrist/frame_001375.jpg", "images/agibotworld/422/713721/cam_right_wrist/frame_001375.jpg", "images/agibotworld/422/713721/cam_high/frame_000964.jpg", "images/agibotworld/422/713721/cam_left_wrist/frame_000964.jpg", "images/agibotworld/422/713721/cam_right_wrist/frame_000964.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pack items for industrial logistics`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pack items for industrial logistics`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -32.5, "hop_absolute_value": 32.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_460_716501-train-bf_000580-af_000338-minus-0000113", "task": "Remove the pillowcase from the clothesline and place it in the basket.", "image": [ "images/agibotworld/460/716501/cam_high/frame_000050.jpg", "images/agibotworld/460/716501/cam_high/frame_001203.jpg", "images/agibotworld/460/716501/cam_high/frame_000580.jpg", "images/agibotworld/460/716501/cam_left_wrist/frame_000580.jpg", "images/agibotworld/460/716501/cam_right_wrist/frame_000580.jpg", "images/agibotworld/460/716501/cam_high/frame_000338.jpg", "images/agibotworld/460/716501/cam_left_wrist/frame_000338.jpg", "images/agibotworld/460/716501/cam_right_wrist/frame_000338.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the pillowcase from the clothesline and place it in the basket.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the pillowcase from the clothesline and place it in the basket.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -63.6, "hop_absolute_value": 63.6, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_575_808324-train-bf_000513-af_000395-minus-0000515", "task": "Store toys", "image": [ "images/agibotworld/575/808324/cam_high/frame_000021.jpg", "images/agibotworld/575/808324/cam_high/frame_000637.jpg", "images/agibotworld/575/808324/cam_high/frame_000513.jpg", "images/agibotworld/575/808324/cam_left_wrist/frame_000513.jpg", "images/agibotworld/575/808324/cam_right_wrist/frame_000513.jpg", "images/agibotworld/575/808324/cam_high/frame_000395.jpg", "images/agibotworld/575/808324/cam_left_wrist/frame_000395.jpg", "images/agibotworld/575/808324/cam_right_wrist/frame_000395.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Store toys`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Store toys`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -23.5, "hop_absolute_value": 23.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_762_925145-train-bf_000581-af_000034-minus-0000014", "task": "Paint the wall", "image": [ "images/agibotworld/762/925145/cam_high/frame_000000.jpg", "images/agibotworld/762/925145/cam_high/frame_001174.jpg", "images/agibotworld/762/925145/cam_high/frame_000581.jpg", "images/agibotworld/762/925145/cam_left_wrist/frame_000581.jpg", "images/agibotworld/762/925145/cam_right_wrist/frame_000581.jpg", "images/agibotworld/762/925145/cam_high/frame_000034.jpg", "images/agibotworld/762/925145/cam_left_wrist/frame_000034.jpg", "images/agibotworld/762/925145/cam_right_wrist/frame_000034.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Paint the wall`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Paint the wall`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -94.4, "hop_absolute_value": 94.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_468_708172-train-bf_001151-af_001106-minus-0000469", "task": "Remove clothes from the drying rack and place them in the basket.", "image": [ "images/agibotworld/468/715715/cam_high/frame_000000.jpg", "images/agibotworld/468/715715/cam_high/frame_000934.jpg", "images/agibotworld/468/708172/cam_high/frame_001151.jpg", "images/agibotworld/468/708172/cam_left_wrist/frame_001151.jpg", "images/agibotworld/468/708172/cam_right_wrist/frame_001151.jpg", "images/agibotworld/468/708172/cam_high/frame_001106.jpg", "images/agibotworld/468/708172/cam_left_wrist/frame_001106.jpg", "images/agibotworld/468/708172/cam_right_wrist/frame_001106.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove clothes from the drying rack and place them in the basket.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove clothes from the drying rack and place them in the basket.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.5, "hop_absolute_value": 4.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_365_688558-train-bf_000832-af_000492-minus-0000751", "task": "Sort personal care products", "image": [ "images/agibotworld/365/688558/cam_high/frame_000036.jpg", "images/agibotworld/365/688558/cam_high/frame_001149.jpg", "images/agibotworld/365/688558/cam_high/frame_000832.jpg", "images/agibotworld/365/688558/cam_left_wrist/frame_000832.jpg", "images/agibotworld/365/688558/cam_right_wrist/frame_000832.jpg", "images/agibotworld/365/688558/cam_high/frame_000492.jpg", "images/agibotworld/365/688558/cam_left_wrist/frame_000492.jpg", "images/agibotworld/365/688558/cam_right_wrist/frame_000492.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort personal care products`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort personal care products`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -36.8, "hop_absolute_value": 36.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_445_795631-train-bf_001194-af_001064-minus-0000606", "task": "Open the fridge to get fruits and vegetables.", "image": [ "images/agibotworld/445/795631/cam_high/frame_000000.jpg", "images/agibotworld/445/795631/cam_high/frame_001546.jpg", "images/agibotworld/445/795631/cam_high/frame_001194.jpg", "images/agibotworld/445/795631/cam_left_wrist/frame_001194.jpg", "images/agibotworld/445/795631/cam_right_wrist/frame_001194.jpg", "images/agibotworld/445/795631/cam_high/frame_001064.jpg", "images/agibotworld/445/795631/cam_left_wrist/frame_001064.jpg", "images/agibotworld/445/795631/cam_right_wrist/frame_001064.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the fridge to get fruits and vegetables.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the fridge to get fruits and vegetables.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.4, "hop_absolute_value": 7.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_598_863074-train-bf_002148-af_000337-minus-0000053", "task": "Pour the tea", "image": [ "images/agibotworld/598/902462/cam_high/frame_000000.jpg", "images/agibotworld/598/902462/cam_high/frame_004395.jpg", "images/agibotworld/598/863074/cam_high/frame_002148.jpg", "images/agibotworld/598/863074/cam_left_wrist/frame_002148.jpg", "images/agibotworld/598/863074/cam_right_wrist/frame_002148.jpg", "images/agibotworld/598/863074/cam_high/frame_000337.jpg", "images/agibotworld/598/863074/cam_left_wrist/frame_000337.jpg", "images/agibotworld/598/863074/cam_right_wrist/frame_000337.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pour the tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pour the tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -84.3, "hop_absolute_value": 84.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_365_688964-train-bf_000933-af_000727-minus-0000656", "task": "Sort personal care products", "image": [ "images/agibotworld/365/688964/cam_high/frame_000000.jpg", "images/agibotworld/365/688964/cam_high/frame_000933.jpg", "images/agibotworld/365/688964/cam_high/frame_000933.jpg", "images/agibotworld/365/688964/cam_left_wrist/frame_000933.jpg", "images/agibotworld/365/688964/cam_right_wrist/frame_000933.jpg", "images/agibotworld/365/688964/cam_high/frame_000727.jpg", "images/agibotworld/365/688964/cam_left_wrist/frame_000727.jpg", "images/agibotworld/365/688964/cam_right_wrist/frame_000727.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Sort personal care products`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Sort personal care products`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -26.7, "hop_absolute_value": 26.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_787_916878-train-bf_000993-af_000378-minus-0000088", "task": "Clear the dining table", "image": [ "images/agibotworld/787/925339/cam_high/frame_000019.jpg", "images/agibotworld/787/925339/cam_high/frame_002703.jpg", "images/agibotworld/787/916878/cam_high/frame_000993.jpg", "images/agibotworld/787/916878/cam_left_wrist/frame_000993.jpg", "images/agibotworld/787/916878/cam_right_wrist/frame_000993.jpg", "images/agibotworld/787/916878/cam_high/frame_000378.jpg", "images/agibotworld/787/916878/cam_left_wrist/frame_000378.jpg", "images/agibotworld/787/916878/cam_right_wrist/frame_000378.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Clear the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Clear the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -78.4, "hop_absolute_value": 78.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_709_891234-train-bf_000418-af_000000-minus-0000001", "task": "Pack the medicines", "image": [ "images/agibotworld/709/891234/cam_high/frame_000000.jpg", "images/agibotworld/709/891234/cam_high/frame_004382.jpg", "images/agibotworld/709/891234/cam_high/frame_000418.jpg", "images/agibotworld/709/891234/cam_left_wrist/frame_000418.jpg", "images/agibotworld/709/891234/cam_right_wrist/frame_000418.jpg", "images/agibotworld/709/891234/cam_high/frame_000000.jpg", "images/agibotworld/709/891234/cam_left_wrist/frame_000000.jpg", "images/agibotworld/709/891234/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pack the medicines`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pack the medicines`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_433_746881-train-bf_003491-af_003152-minus-0000905", "task": "Make instant coffee", "image": [ "images/agibotworld/433/726834/cam_high/frame_000000.jpg", "images/agibotworld/433/726834/cam_high/frame_003765.jpg", "images/agibotworld/433/746881/cam_high/frame_003491.jpg", "images/agibotworld/433/746881/cam_left_wrist/frame_003491.jpg", "images/agibotworld/433/746881/cam_right_wrist/frame_003491.jpg", "images/agibotworld/433/746881/cam_high/frame_003152.jpg", "images/agibotworld/433/746881/cam_left_wrist/frame_003152.jpg", "images/agibotworld/433/746881/cam_right_wrist/frame_003152.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make instant coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make instant coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.9, "hop_absolute_value": 11.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_533_873979-train-bf_000705-af_000473-minus-0000180", "task": "Open the curtains.", "image": [ "images/agibotworld/533/876110/cam_high/frame_000102.jpg", "images/agibotworld/533/876110/cam_high/frame_000768.jpg", "images/agibotworld/533/873979/cam_high/frame_000705.jpg", "images/agibotworld/533/873979/cam_left_wrist/frame_000705.jpg", "images/agibotworld/533/873979/cam_right_wrist/frame_000705.jpg", "images/agibotworld/533/873979/cam_high/frame_000473.jpg", "images/agibotworld/533/873979/cam_left_wrist/frame_000473.jpg", "images/agibotworld/533/873979/cam_right_wrist/frame_000473.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the curtains.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the curtains.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -42.1, "hop_absolute_value": 42.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_392_680342-train-bf_002696-af_002601-minus-0000300", "task": "Brush water bottle", "image": [ "images/agibotworld/392/651277/cam_high/frame_000025.jpg", "images/agibotworld/392/651277/cam_high/frame_003057.jpg", "images/agibotworld/392/680342/cam_high/frame_002696.jpg", "images/agibotworld/392/680342/cam_left_wrist/frame_002696.jpg", "images/agibotworld/392/680342/cam_right_wrist/frame_002696.jpg", "images/agibotworld/392/680342/cam_high/frame_002601.jpg", "images/agibotworld/392/680342/cam_left_wrist/frame_002601.jpg", "images/agibotworld/392/680342/cam_right_wrist/frame_002601.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Brush water bottle`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Brush water bottle`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.2, "hop_absolute_value": 3.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_580_868347-train-bf_002623-af_002569-minus-0000713", "task": "Make juice", "image": [ "images/agibotworld/580/868347/cam_high/frame_000000.jpg", "images/agibotworld/580/868347/cam_high/frame_006708.jpg", "images/agibotworld/580/868347/cam_high/frame_002623.jpg", "images/agibotworld/580/868347/cam_left_wrist/frame_002623.jpg", "images/agibotworld/580/868347/cam_right_wrist/frame_002623.jpg", "images/agibotworld/580/868347/cam_high/frame_002569.jpg", "images/agibotworld/580/868347/cam_left_wrist/frame_002569.jpg", "images/agibotworld/580/868347/cam_right_wrist/frame_002569.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make juice`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make juice`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.4, "hop_absolute_value": 1.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_544_817708-train-bf_001406-af_001338-minus-0000748", "task": "Wash the pot and spatula", "image": [ "images/agibotworld/544/817540/cam_high/frame_000039.jpg", "images/agibotworld/544/817540/cam_high/frame_002255.jpg", "images/agibotworld/544/817708/cam_high/frame_001406.jpg", "images/agibotworld/544/817708/cam_left_wrist/frame_001406.jpg", "images/agibotworld/544/817708/cam_right_wrist/frame_001406.jpg", "images/agibotworld/544/817708/cam_high/frame_001338.jpg", "images/agibotworld/544/817708/cam_left_wrist/frame_001338.jpg", "images/agibotworld/544/817708/cam_right_wrist/frame_001338.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wash the pot and spatula`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wash the pot and spatula`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.9, "hop_absolute_value": 3.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_707_870986-train-bf_001475-af_001233-minus-0000671", "task": "Twist the bottle cap", "image": [ "images/agibotworld/707/870786/cam_high/frame_000000.jpg", "images/agibotworld/707/870786/cam_high/frame_002312.jpg", "images/agibotworld/707/870986/cam_high/frame_001475.jpg", "images/agibotworld/707/870986/cam_left_wrist/frame_001475.jpg", "images/agibotworld/707/870986/cam_right_wrist/frame_001475.jpg", "images/agibotworld/707/870986/cam_high/frame_001233.jpg", "images/agibotworld/707/870986/cam_left_wrist/frame_001233.jpg", "images/agibotworld/707/870986/cam_right_wrist/frame_001233.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -15.8, "hop_absolute_value": 15.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_548_834508-train-bf_000386-af_000360-minus-0000286", "task": "Separate dark and light clothes", "image": [ "images/agibotworld/548/834029/cam_high/frame_000000.jpg", "images/agibotworld/548/834029/cam_high/frame_001546.jpg", "images/agibotworld/548/834508/cam_high/frame_000386.jpg", "images/agibotworld/548/834508/cam_left_wrist/frame_000386.jpg", "images/agibotworld/548/834508/cam_right_wrist/frame_000386.jpg", "images/agibotworld/548/834508/cam_high/frame_000360.jpg", "images/agibotworld/548/834508/cam_left_wrist/frame_000360.jpg", "images/agibotworld/548/834508/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Separate dark and light clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Separate dark and light clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -8.3, "hop_absolute_value": 8.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_527_789530-train-bf_001653-af_001630-minus-0000867", "task": "Arrange sofa", "image": [ "images/agibotworld/527/784377/cam_high/frame_000000.jpg", "images/agibotworld/527/784377/cam_high/frame_005225.jpg", "images/agibotworld/527/789530/cam_high/frame_001653.jpg", "images/agibotworld/527/789530/cam_left_wrist/frame_001653.jpg", "images/agibotworld/527/789530/cam_right_wrist/frame_001653.jpg", "images/agibotworld/527/789530/cam_high/frame_001630.jpg", "images/agibotworld/527/789530/cam_left_wrist/frame_001630.jpg", "images/agibotworld/527/789530/cam_right_wrist/frame_001630.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Arrange sofa`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Arrange sofa`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.6, "hop_absolute_value": 3.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_589_852398-train-bf_001154-af_001121-minus-0000805", "task": "Shred vegetables with a slicer.", "image": [ "images/agibotworld/589/854912/cam_high/frame_000000.jpg", "images/agibotworld/589/854912/cam_high/frame_002032.jpg", "images/agibotworld/589/852398/cam_high/frame_001154.jpg", "images/agibotworld/589/852398/cam_left_wrist/frame_001154.jpg", "images/agibotworld/589/852398/cam_right_wrist/frame_001154.jpg", "images/agibotworld/589/852398/cam_high/frame_001121.jpg", "images/agibotworld/589/852398/cam_left_wrist/frame_001121.jpg", "images/agibotworld/589/852398/cam_right_wrist/frame_001121.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Shred vegetables with a slicer.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Shred vegetables with a slicer.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.6, "hop_absolute_value": 3.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agibot_532_792567-train-bf_000031-af_000000-minus-0000000", "task": "Packing permanent magnet ingot", "image": [ "images/agibotworld/532/788703/cam_high/frame_000000.jpg", "images/agibotworld/532/788703/cam_high/frame_000920.jpg", "images/agibotworld/532/792567/cam_high/frame_000031.jpg", "images/agibotworld/532/792567/cam_left_wrist/frame_000031.jpg", "images/agibotworld/532/792567/cam_right_wrist/frame_000031.jpg", "images/agibotworld/532/792567/cam_high/frame_000000.jpg", "images/agibotworld/532/792567/cam_left_wrist/frame_000000.jpg", "images/agibotworld/532/792567/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Packing permanent magnet ingot`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Packing permanent magnet ingot`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agibotworld" }, { "id": "agibot_434_831844-train-bf_000899-af_000529-minus-0000168", "task": "Remove capsules from the coffee machine.", "image": [ "images/agibotworld/434/831844/cam_high/frame_000000.jpg", "images/agibotworld/434/831844/cam_high/frame_001496.jpg", "images/agibotworld/434/831844/cam_high/frame_000899.jpg", "images/agibotworld/434/831844/cam_left_wrist/frame_000899.jpg", "images/agibotworld/434/831844/cam_right_wrist/frame_000899.jpg", "images/agibotworld/434/831844/cam_high/frame_000529.jpg", "images/agibotworld/434/831844/cam_left_wrist/frame_000529.jpg", "images/agibotworld/434/831844/cam_right_wrist/frame_000529.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove capsules from the coffee machine.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove capsules from the coffee machine.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -55.0, "hop_absolute_value": 55.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agibotworld" }, { "id": "agibot_580_868478-train-bf_003543-af_003471-minus-0000748", "task": "Make juice", "image": [ "images/agibotworld/580/862866/cam_high/frame_000000.jpg", "images/agibotworld/580/862866/cam_high/frame_007036.jpg", "images/agibotworld/580/868478/cam_high/frame_003543.jpg", "images/agibotworld/580/868478/cam_left_wrist/frame_003543.jpg", "images/agibotworld/580/868478/cam_right_wrist/frame_003543.jpg", "images/agibotworld/580/868478/cam_high/frame_003471.jpg", "images/agibotworld/580/868478/cam_left_wrist/frame_003471.jpg", "images/agibotworld/580/868478/cam_right_wrist/frame_003471.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make juice`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make juice`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.2, "hop_absolute_value": 1.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agibotworld" }, { "id": "agilex_newdragon_clean_the_desk_episode_17-train-bf_000720-af_001215-plus-0000417", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_23/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_high/frame_001845.jpg", "images/agilex_newdragon/clean_the_desk/episode_17/cam_high/frame_000720.jpg", "images/agilex_newdragon/clean_the_desk/episode_17/cam_left_wrist/frame_000720.jpg", "images/agilex_newdragon/clean_the_desk/episode_17/cam_right_wrist/frame_000720.jpg", "images/agilex_newdragon/clean_the_desk/episode_17/cam_high/frame_001215.jpg", "images/agilex_newdragon/clean_the_desk/episode_17/cam_left_wrist/frame_001215.jpg", "images/agilex_newdragon/clean_the_desk/episode_17/cam_right_wrist/frame_001215.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 49.3, "hop_absolute_value": 49.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_2-train-bf_000270-af_000360-plus-0000327", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_23/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_23/cam_high/frame_000660.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_2/cam_high/frame_000270.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_2/cam_left_wrist/frame_000270.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_2/cam_right_wrist/frame_000270.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_2/cam_high/frame_000360.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_2/cam_left_wrist/frame_000360.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_2/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 30.0, "hop_absolute_value": 30.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_11-train-bf_000165-af_000255-plus-0000749", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_3/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_3/cam_high/frame_000527.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_11/cam_high/frame_000165.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_11/cam_left_wrist/frame_000165.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_11/cam_right_wrist/frame_000165.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_11/cam_high/frame_000255.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_11/cam_left_wrist/frame_000255.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_11/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 30.0, "hop_absolute_value": 30.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_12-train-bf_000135-af_000210-plus-0000653", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_12/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_12/cam_high/frame_000445.jpg", "images/agilex_newdragon/plug_insertion/episode_12/cam_high/frame_000135.jpg", "images/agilex_newdragon/plug_insertion/episode_12/cam_left_wrist/frame_000135.jpg", "images/agilex_newdragon/plug_insertion/episode_12/cam_right_wrist/frame_000135.jpg", "images/agilex_newdragon/plug_insertion/episode_12/cam_high/frame_000210.jpg", "images/agilex_newdragon/plug_insertion/episode_12/cam_left_wrist/frame_000210.jpg", "images/agilex_newdragon/plug_insertion/episode_12/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_clothes_episode_15-train-bf_000315-af_000924-plus-0000593", "task": "fold clothes", "image": [ "images/agilex_newdragon/fold_clothes/episode_17/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_clothes/episode_17/cam_high/frame_000901.jpg", "images/agilex_newdragon/fold_clothes/episode_15/cam_high/frame_000315.jpg", "images/agilex_newdragon/fold_clothes/episode_15/cam_left_wrist/frame_000315.jpg", "images/agilex_newdragon/fold_clothes/episode_15/cam_right_wrist/frame_000315.jpg", "images/agilex_newdragon/fold_clothes/episode_15/cam_high/frame_000924.jpg", "images/agilex_newdragon/fold_clothes/episode_15/cam_left_wrist/frame_000924.jpg", "images/agilex_newdragon/fold_clothes/episode_15/cam_right_wrist/frame_000924.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_6-train-bf_000375-af_000390-plus-0000666", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_19/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_19/cam_high/frame_001456.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_high/frame_000375.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_left_wrist/frame_000375.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_right_wrist/frame_000375.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_high/frame_000390.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_left_wrist/frame_000390.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.0, "hop_absolute_value": 2.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_20-train-bf_000855-af_000990-plus-0000394", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_1/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_1/cam_high/frame_001461.jpg", "images/agilex_newdragon/pour_water/episode_20/cam_high/frame_000855.jpg", "images/agilex_newdragon/pour_water/episode_20/cam_left_wrist/frame_000855.jpg", "images/agilex_newdragon/pour_water/episode_20/cam_right_wrist/frame_000855.jpg", "images/agilex_newdragon/pour_water/episode_20/cam_high/frame_000990.jpg", "images/agilex_newdragon/pour_water/episode_20/cam_left_wrist/frame_000990.jpg", "images/agilex_newdragon/pour_water/episode_20/cam_right_wrist/frame_000990.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 45.0, "hop_absolute_value": 45.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_17-train-bf_000060-af_000270-plus-0000820", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_17/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_17/cam_high/frame_000465.jpg", "images/agilex_newdragon/plug_insertion/episode_17/cam_high/frame_000060.jpg", "images/agilex_newdragon/plug_insertion/episode_17/cam_left_wrist/frame_000060.jpg", "images/agilex_newdragon/plug_insertion/episode_17/cam_right_wrist/frame_000060.jpg", "images/agilex_newdragon/plug_insertion/episode_17/cam_high/frame_000270.jpg", "images/agilex_newdragon/plug_insertion/episode_17/cam_left_wrist/frame_000270.jpg", "images/agilex_newdragon/plug_insertion/episode_17/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 51.9, "hop_absolute_value": 51.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_23-train-bf_000075-af_000345-plus-0000384", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_8/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_8/cam_high/frame_000624.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_23/cam_high/frame_000075.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_23/cam_left_wrist/frame_000075.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_23/cam_right_wrist/frame_000075.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_23/cam_high/frame_000345.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_23/cam_left_wrist/frame_000345.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_23/cam_right_wrist/frame_000345.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 46.2, "hop_absolute_value": 46.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_23-train-bf_000135-af_000150-plus-0000554", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_9/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_high/frame_000563.jpg", "images/agilex_newdragon/glasses_case/episode_23/cam_high/frame_000135.jpg", "images/agilex_newdragon/glasses_case/episode_23/cam_left_wrist/frame_000135.jpg", "images/agilex_newdragon/glasses_case/episode_23/cam_right_wrist/frame_000135.jpg", "images/agilex_newdragon/glasses_case/episode_23/cam_high/frame_000150.jpg", "images/agilex_newdragon/glasses_case/episode_23/cam_left_wrist/frame_000150.jpg", "images/agilex_newdragon/glasses_case/episode_23/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.0, "hop_absolute_value": 5.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_13-train-bf_000360-af_000435-plus-0000424", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_19/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_19/cam_high/frame_000706.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_high/frame_000360.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_left_wrist/frame_000360.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_right_wrist/frame_000360.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_high/frame_000435.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_left_wrist/frame_000435.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_right_wrist/frame_000435.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_18-train-bf_000375-af_000585-plus-0000382", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_13/cam_high/frame_001246.jpg", "images/agilex_newdragon/glasses_storage/episode_18/cam_high/frame_000375.jpg", "images/agilex_newdragon/glasses_storage/episode_18/cam_left_wrist/frame_000375.jpg", "images/agilex_newdragon/glasses_storage/episode_18/cam_right_wrist/frame_000375.jpg", "images/agilex_newdragon/glasses_storage/episode_18/cam_high/frame_000585.jpg", "images/agilex_newdragon/glasses_storage/episode_18/cam_left_wrist/frame_000585.jpg", "images/agilex_newdragon/glasses_storage/episode_18/cam_right_wrist/frame_000585.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 46.7, "hop_absolute_value": 46.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_16-train-bf_000075-af_000315-plus-0000716", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_3/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_3/cam_high/frame_000388.jpg", "images/agilex_newdragon/glasses_case/episode_16/cam_high/frame_000075.jpg", "images/agilex_newdragon/glasses_case/episode_16/cam_left_wrist/frame_000075.jpg", "images/agilex_newdragon/glasses_case/episode_16/cam_right_wrist/frame_000075.jpg", "images/agilex_newdragon/glasses_case/episode_16/cam_high/frame_000315.jpg", "images/agilex_newdragon/glasses_case/episode_16/cam_left_wrist/frame_000315.jpg", "images/agilex_newdragon/glasses_case/episode_16/cam_right_wrist/frame_000315.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 72.7, "hop_absolute_value": 72.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_15-train-bf_000045-af_000210-plus-0000330", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_6/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_6/cam_high/frame_000473.jpg", "images/agilex_newdragon/plug_insertion/episode_15/cam_high/frame_000045.jpg", "images/agilex_newdragon/plug_insertion/episode_15/cam_left_wrist/frame_000045.jpg", "images/agilex_newdragon/plug_insertion/episode_15/cam_right_wrist/frame_000045.jpg", "images/agilex_newdragon/plug_insertion/episode_15/cam_high/frame_000210.jpg", "images/agilex_newdragon/plug_insertion/episode_15/cam_left_wrist/frame_000210.jpg", "images/agilex_newdragon/plug_insertion/episode_15/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_7-train-bf_000000-af_000330-plus-0000409", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_21/cam_high/frame_000538.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_7/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_7/cam_left_wrist/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_7/cam_right_wrist/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_7/cam_high/frame_000330.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_7/cam_left_wrist/frame_000330.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_7/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.9, "hop_absolute_value": 57.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_24-train-bf_000120-af_000135-plus-0000791", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_high/frame_001698.jpg", "images/agilex_newdragon/food_packaging/episode_24/cam_high/frame_000120.jpg", "images/agilex_newdragon/food_packaging/episode_24/cam_left_wrist/frame_000120.jpg", "images/agilex_newdragon/food_packaging/episode_24/cam_right_wrist/frame_000120.jpg", "images/agilex_newdragon/food_packaging/episode_24/cam_high/frame_000135.jpg", "images/agilex_newdragon/food_packaging/episode_24/cam_left_wrist/frame_000135.jpg", "images/agilex_newdragon/food_packaging/episode_24/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.1, "hop_absolute_value": 1.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_12-train-bf_001275-af_001530-plus-0000422", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_1/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_1/cam_high/frame_001493.jpg", "images/agilex_newdragon/clean_the_desk/episode_12/cam_high/frame_001275.jpg", "images/agilex_newdragon/clean_the_desk/episode_12/cam_left_wrist/frame_001275.jpg", "images/agilex_newdragon/clean_the_desk/episode_12/cam_right_wrist/frame_001275.jpg", "images/agilex_newdragon/clean_the_desk/episode_12/cam_high/frame_001530.jpg", "images/agilex_newdragon/clean_the_desk/episode_12/cam_left_wrist/frame_001530.jpg", "images/agilex_newdragon/clean_the_desk/episode_12/cam_right_wrist/frame_001530.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 58.6, "hop_absolute_value": 58.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_9-train-bf_000135-af_000210-plus-0000767", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_0/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_0/cam_high/frame_000383.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_high/frame_000135.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_left_wrist/frame_000135.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_right_wrist/frame_000135.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_high/frame_000210.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_left_wrist/frame_000210.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 17.9, "hop_absolute_value": 17.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_22-train-bf_000495-af_000570-plus-0000373", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_9/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_9/cam_high/frame_000525.jpg", "images/agilex_newdragon/unplug/episode_22/cam_high/frame_000495.jpg", "images/agilex_newdragon/unplug/episode_22/cam_left_wrist/frame_000495.jpg", "images/agilex_newdragon/unplug/episode_22/cam_right_wrist/frame_000495.jpg", "images/agilex_newdragon/unplug/episode_22/cam_high/frame_000570.jpg", "images/agilex_newdragon/unplug/episode_22/cam_left_wrist/frame_000570.jpg", "images/agilex_newdragon/unplug/episode_22/cam_right_wrist/frame_000570.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_14-train-bf_000225-af_000255-plus-0000673", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_21/cam_high/frame_000538.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_14/cam_high/frame_000225.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_14/cam_left_wrist/frame_000225.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_14/cam_right_wrist/frame_000225.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_14/cam_high/frame_000255.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_14/cam_left_wrist/frame_000255.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_14/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 7.4, "hop_absolute_value": 7.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_4-train-bf_000195-af_000270-plus-0000291", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_4/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_high/frame_000413.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_high/frame_000195.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_left_wrist/frame_000195.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_right_wrist/frame_000195.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_high/frame_000270.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_left_wrist/frame_000270.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 35.7, "hop_absolute_value": 35.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_5-train-bf_000315-af_000465-plus-0000352", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_0/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_0/cam_high/frame_000616.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_5/cam_high/frame_000315.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_5/cam_left_wrist/frame_000315.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_5/cam_right_wrist/frame_000315.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_5/cam_high/frame_000465.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_5/cam_left_wrist/frame_000465.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_5/cam_right_wrist/frame_000465.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 43.5, "hop_absolute_value": 43.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_13-train-bf_000000-af_000390-plus-0000463", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_13/cam_high/frame_000562.jpg", "images/agilex_newdragon/plug_insertion/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_13/cam_left_wrist/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_13/cam_right_wrist/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_13/cam_high/frame_000390.jpg", "images/agilex_newdragon/plug_insertion/episode_13/cam_left_wrist/frame_000390.jpg", "images/agilex_newdragon/plug_insertion/episode_13/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.3, "hop_absolute_value": 70.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_9-train-bf_001095-af_001170-plus-0000449", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_9/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_9/cam_high/frame_001216.jpg", "images/agilex_newdragon/pour_water/episode_9/cam_high/frame_001095.jpg", "images/agilex_newdragon/pour_water/episode_9/cam_left_wrist/frame_001095.jpg", "images/agilex_newdragon/pour_water/episode_9/cam_right_wrist/frame_001095.jpg", "images/agilex_newdragon/pour_water/episode_9/cam_high/frame_001170.jpg", "images/agilex_newdragon/pour_water/episode_9/cam_left_wrist/frame_001170.jpg", "images/agilex_newdragon/pour_water/episode_9/cam_right_wrist/frame_001170.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_10-train-bf_001290-af_001380-plus-0000452", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_24/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_24/cam_high/frame_001192.jpg", "images/agilex_newdragon/pour_coffee/episode_10/cam_high/frame_001290.jpg", "images/agilex_newdragon/pour_coffee/episode_10/cam_left_wrist/frame_001290.jpg", "images/agilex_newdragon/pour_coffee/episode_10/cam_right_wrist/frame_001290.jpg", "images/agilex_newdragon/pour_coffee/episode_10/cam_high/frame_001380.jpg", "images/agilex_newdragon/pour_coffee/episode_10/cam_left_wrist/frame_001380.jpg", "images/agilex_newdragon/pour_coffee/episode_10/cam_right_wrist/frame_001380.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_2-train-bf_000015-af_000315-plus-0000401", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_15/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_15/cam_high/frame_000549.jpg", "images/agilex_newdragon/plug_insertion/episode_2/cam_high/frame_000015.jpg", "images/agilex_newdragon/plug_insertion/episode_2/cam_left_wrist/frame_000015.jpg", "images/agilex_newdragon/plug_insertion/episode_2/cam_right_wrist/frame_000015.jpg", "images/agilex_newdragon/plug_insertion/episode_2/cam_high/frame_000315.jpg", "images/agilex_newdragon/plug_insertion/episode_2/cam_left_wrist/frame_000315.jpg", "images/agilex_newdragon/plug_insertion/episode_2/cam_right_wrist/frame_000315.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 58.8, "hop_absolute_value": 58.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_5-train-bf_000570-af_000675-plus-0000519", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_24/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_24/cam_high/frame_000992.jpg", "images/agilex_newdragon/glasses_storage/episode_5/cam_high/frame_000570.jpg", "images/agilex_newdragon/glasses_storage/episode_5/cam_left_wrist/frame_000570.jpg", "images/agilex_newdragon/glasses_storage/episode_5/cam_right_wrist/frame_000570.jpg", "images/agilex_newdragon/glasses_storage/episode_5/cam_high/frame_000675.jpg", "images/agilex_newdragon/glasses_storage/episode_5/cam_left_wrist/frame_000675.jpg", "images/agilex_newdragon/glasses_storage/episode_5/cam_right_wrist/frame_000675.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_0-train-bf_000105-af_000465-plus-0000546", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_24/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_24/cam_high/frame_000764.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_0/cam_high/frame_000105.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_0/cam_left_wrist/frame_000105.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_0/cam_right_wrist/frame_000105.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_0/cam_high/frame_000465.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_0/cam_left_wrist/frame_000465.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_0/cam_right_wrist/frame_000465.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 92.3, "hop_absolute_value": 92.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_10-train-bf_000165-af_000615-plus-0000540", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_high/frame_000667.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_high/frame_000165.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_left_wrist/frame_000165.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_right_wrist/frame_000165.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_high/frame_000615.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_left_wrist/frame_000615.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_right_wrist/frame_000615.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 90.9, "hop_absolute_value": 90.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_13-train-bf_000255-af_000405-plus-0000903", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_high/frame_000667.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_high/frame_000255.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_left_wrist/frame_000255.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_right_wrist/frame_000255.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_high/frame_000405.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_left_wrist/frame_000405.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_13-train-bf_000870-af_000900-plus-0000312", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_13/cam_high/frame_001246.jpg", "images/agilex_newdragon/glasses_storage/episode_13/cam_high/frame_000870.jpg", "images/agilex_newdragon/glasses_storage/episode_13/cam_left_wrist/frame_000870.jpg", "images/agilex_newdragon/glasses_storage/episode_13/cam_right_wrist/frame_000870.jpg", "images/agilex_newdragon/glasses_storage/episode_13/cam_high/frame_000900.jpg", "images/agilex_newdragon/glasses_storage/episode_13/cam_left_wrist/frame_000900.jpg", "images/agilex_newdragon/glasses_storage/episode_13/cam_right_wrist/frame_000900.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 8.0, "hop_absolute_value": 8.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_2-train-bf_000075-af_000600-plus-0000455", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_4/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_4/cam_high/frame_001061.jpg", "images/agilex_newdragon/food_packaging/episode_2/cam_high/frame_000075.jpg", "images/agilex_newdragon/food_packaging/episode_2/cam_left_wrist/frame_000075.jpg", "images/agilex_newdragon/food_packaging/episode_2/cam_right_wrist/frame_000075.jpg", "images/agilex_newdragon/food_packaging/episode_2/cam_high/frame_000600.jpg", "images/agilex_newdragon/food_packaging/episode_2/cam_left_wrist/frame_000600.jpg", "images/agilex_newdragon/food_packaging/episode_2/cam_right_wrist/frame_000600.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_10-train-bf_000195-af_000240-plus-0000869", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_0/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_0/cam_high/frame_000987.jpg", "images/agilex_newdragon/clean_the_desk/episode_10/cam_high/frame_000195.jpg", "images/agilex_newdragon/clean_the_desk/episode_10/cam_left_wrist/frame_000195.jpg", "images/agilex_newdragon/clean_the_desk/episode_10/cam_right_wrist/frame_000195.jpg", "images/agilex_newdragon/clean_the_desk/episode_10/cam_high/frame_000240.jpg", "images/agilex_newdragon/clean_the_desk/episode_10/cam_left_wrist/frame_000240.jpg", "images/agilex_newdragon/clean_the_desk/episode_10/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.4, "hop_absolute_value": 3.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_0-train-bf_000435-af_000465-plus-0000823", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_12/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_12/cam_high/frame_001979.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_high/frame_000435.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_left_wrist/frame_000435.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_right_wrist/frame_000435.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_high/frame_000465.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_left_wrist/frame_000465.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_right_wrist/frame_000465.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.9, "hop_absolute_value": 1.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_10-train-bf_000135-af_000270-plus-0000330", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_14/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_14/cam_high/frame_000508.jpg", "images/agilex_newdragon/plug_insertion/episode_10/cam_high/frame_000135.jpg", "images/agilex_newdragon/plug_insertion/episode_10/cam_left_wrist/frame_000135.jpg", "images/agilex_newdragon/plug_insertion/episode_10/cam_right_wrist/frame_000135.jpg", "images/agilex_newdragon/plug_insertion/episode_10/cam_high/frame_000270.jpg", "images/agilex_newdragon/plug_insertion/episode_10/cam_left_wrist/frame_000270.jpg", "images/agilex_newdragon/plug_insertion/episode_10/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 31.0, "hop_absolute_value": 31.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_14-train-bf_000135-af_000165-plus-0000651", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_8/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_8/cam_high/frame_002314.jpg", "images/agilex_newdragon/fold_the_clothes/episode_14/cam_high/frame_000135.jpg", "images/agilex_newdragon/fold_the_clothes/episode_14/cam_left_wrist/frame_000135.jpg", "images/agilex_newdragon/fold_the_clothes/episode_14/cam_right_wrist/frame_000135.jpg", "images/agilex_newdragon/fold_the_clothes/episode_14/cam_high/frame_000165.jpg", "images/agilex_newdragon/fold_the_clothes/episode_14/cam_left_wrist/frame_000165.jpg", "images/agilex_newdragon/fold_the_clothes/episode_14/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.9, "hop_absolute_value": 1.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_19-train-bf_000255-af_000390-plus-0000406", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_20/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_20/cam_high/frame_000387.jpg", "images/agilex_newdragon/glasses_case/episode_19/cam_high/frame_000255.jpg", "images/agilex_newdragon/glasses_case/episode_19/cam_left_wrist/frame_000255.jpg", "images/agilex_newdragon/glasses_case/episode_19/cam_right_wrist/frame_000255.jpg", "images/agilex_newdragon/glasses_case/episode_19/cam_high/frame_000390.jpg", "images/agilex_newdragon/glasses_case/episode_19/cam_left_wrist/frame_000390.jpg", "images/agilex_newdragon/glasses_case/episode_19/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 64.3, "hop_absolute_value": 64.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_clothes_episode_24-train-bf_000195-af_000270-plus-0000895", "task": "fold clothes", "image": [ "images/agilex_newdragon/fold_clothes/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_clothes/episode_22/cam_high/frame_000896.jpg", "images/agilex_newdragon/fold_clothes/episode_24/cam_high/frame_000195.jpg", "images/agilex_newdragon/fold_clothes/episode_24/cam_left_wrist/frame_000195.jpg", "images/agilex_newdragon/fold_clothes/episode_24/cam_right_wrist/frame_000195.jpg", "images/agilex_newdragon/fold_clothes/episode_24/cam_high/frame_000270.jpg", "images/agilex_newdragon/fold_clothes/episode_24/cam_left_wrist/frame_000270.jpg", "images/agilex_newdragon/fold_clothes/episode_24/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.4, "hop_absolute_value": 10.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_0-train-bf_000660-af_001230-plus-0000519", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_18/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_18/cam_high/frame_001415.jpg", "images/agilex_newdragon/pour_water/episode_0/cam_high/frame_000660.jpg", "images/agilex_newdragon/pour_water/episode_0/cam_left_wrist/frame_000660.jpg", "images/agilex_newdragon/pour_water/episode_0/cam_right_wrist/frame_000660.jpg", "images/agilex_newdragon/pour_water/episode_0/cam_high/frame_001230.jpg", "images/agilex_newdragon/pour_water/episode_0/cam_left_wrist/frame_001230.jpg", "images/agilex_newdragon/pour_water/episode_0/cam_right_wrist/frame_001230.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 77.6, "hop_absolute_value": 77.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_9-train-bf_000315-af_001290-plus-0000529", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_19/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_19/cam_high/frame_001127.jpg", "images/agilex_newdragon/pour_coffee/episode_9/cam_high/frame_000315.jpg", "images/agilex_newdragon/pour_coffee/episode_9/cam_left_wrist/frame_000315.jpg", "images/agilex_newdragon/pour_coffee/episode_9/cam_right_wrist/frame_000315.jpg", "images/agilex_newdragon/pour_coffee/episode_9/cam_high/frame_001290.jpg", "images/agilex_newdragon/pour_coffee/episode_9/cam_left_wrist/frame_001290.jpg", "images/agilex_newdragon/pour_coffee/episode_9/cam_right_wrist/frame_001290.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 82.3, "hop_absolute_value": 82.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_16-train-bf_000075-af_000300-plus-0000380", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_16/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_16/cam_high/frame_000468.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_16/cam_high/frame_000075.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_16/cam_left_wrist/frame_000075.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_16/cam_right_wrist/frame_000075.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_16/cam_high/frame_000300.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_16/cam_left_wrist/frame_000300.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_16/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.7, "hop_absolute_value": 57.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_17-train-bf_000510-af_001050-plus-0000418", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_0/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_0/cam_high/frame_000999.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_high/frame_000510.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_left_wrist/frame_000510.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_right_wrist/frame_000510.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_high/frame_001050.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_left_wrist/frame_001050.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_right_wrist/frame_001050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.7, "hop_absolute_value": 50.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_11-train-bf_000540-af_001305-plus-0000532", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_23/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_high/frame_001173.jpg", "images/agilex_newdragon/pour_coffee/episode_11/cam_high/frame_000540.jpg", "images/agilex_newdragon/pour_coffee/episode_11/cam_left_wrist/frame_000540.jpg", "images/agilex_newdragon/pour_coffee/episode_11/cam_right_wrist/frame_000540.jpg", "images/agilex_newdragon/pour_coffee/episode_11/cam_high/frame_001305.jpg", "images/agilex_newdragon/pour_coffee/episode_11/cam_left_wrist/frame_001305.jpg", "images/agilex_newdragon/pour_coffee/episode_11/cam_right_wrist/frame_001305.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 78.5, "hop_absolute_value": 78.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_23-train-bf_000105-af_001065-plus-0000566", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_23/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_high/frame_001173.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_high/frame_000105.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_left_wrist/frame_000105.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_right_wrist/frame_000105.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_high/frame_001065.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_left_wrist/frame_001065.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_right_wrist/frame_001065.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 90.1, "hop_absolute_value": 90.1, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_18-train-bf_000000-af_000285-plus-0000878", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_1/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_1/cam_high/frame_000478.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_left_wrist/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_right_wrist/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_high/frame_000285.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_left_wrist/frame_000285.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 61.3, "hop_absolute_value": 61.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_22-train-bf_000975-af_001425-plus-0000466", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_high/frame_001698.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_high/frame_000975.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_left_wrist/frame_000975.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_right_wrist/frame_000975.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_high/frame_001425.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_left_wrist/frame_001425.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_right_wrist/frame_001425.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_1-train-bf_000090-af_000510-plus-0000449", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_21/cam_high/frame_000731.jpg", "images/agilex_newdragon/unplug/episode_1/cam_high/frame_000090.jpg", "images/agilex_newdragon/unplug/episode_1/cam_left_wrist/frame_000090.jpg", "images/agilex_newdragon/unplug/episode_1/cam_right_wrist/frame_000090.jpg", "images/agilex_newdragon/unplug/episode_1/cam_high/frame_000510.jpg", "images/agilex_newdragon/unplug/episode_1/cam_left_wrist/frame_000510.jpg", "images/agilex_newdragon/unplug/episode_1/cam_right_wrist/frame_000510.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 65.1, "hop_absolute_value": 65.1, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_15-train-bf_000840-af_000870-plus-0000733", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_13/cam_high/frame_001468.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_high/frame_000840.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_left_wrist/frame_000840.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_right_wrist/frame_000840.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_high/frame_000870.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_left_wrist/frame_000870.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_right_wrist/frame_000870.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.8, "hop_absolute_value": 3.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_19-train-bf_000255-af_000465-plus-0000389", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_4/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_4/cam_high/frame_000487.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_19/cam_high/frame_000255.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_19/cam_left_wrist/frame_000255.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_19/cam_right_wrist/frame_000255.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_19/cam_high/frame_000465.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_19/cam_left_wrist/frame_000465.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_19/cam_right_wrist/frame_000465.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 46.7, "hop_absolute_value": 46.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_22-train-bf_000255-af_000525-plus-0000540", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_18/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_18/cam_high/frame_000488.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_high/frame_000255.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_left_wrist/frame_000255.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_right_wrist/frame_000255.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_high/frame_000525.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_left_wrist/frame_000525.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_right_wrist/frame_000525.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 94.7, "hop_absolute_value": 94.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_24-train-bf_000135-af_000315-plus-0000355", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_18/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_18/cam_high/frame_001159.jpg", "images/agilex_newdragon/unplug/episode_24/cam_high/frame_000135.jpg", "images/agilex_newdragon/unplug/episode_24/cam_left_wrist/frame_000135.jpg", "images/agilex_newdragon/unplug/episode_24/cam_right_wrist/frame_000135.jpg", "images/agilex_newdragon/unplug/episode_24/cam_high/frame_000315.jpg", "images/agilex_newdragon/unplug/episode_24/cam_left_wrist/frame_000315.jpg", "images/agilex_newdragon/unplug/episode_24/cam_right_wrist/frame_000315.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_23-train-bf_000015-af_000180-plus-0000334", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_high/frame_000558.jpg", "images/agilex_newdragon/pour_water/episode_23/cam_high/frame_000015.jpg", "images/agilex_newdragon/pour_water/episode_23/cam_left_wrist/frame_000015.jpg", "images/agilex_newdragon/pour_water/episode_23/cam_right_wrist/frame_000015.jpg", "images/agilex_newdragon/pour_water/episode_23/cam_high/frame_000180.jpg", "images/agilex_newdragon/pour_water/episode_23/cam_left_wrist/frame_000180.jpg", "images/agilex_newdragon/pour_water/episode_23/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_1-train-bf_000435-af_000450-plus-0000617", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_5/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_5/cam_high/frame_000692.jpg", "images/agilex_newdragon/glasses_storage/episode_1/cam_high/frame_000435.jpg", "images/agilex_newdragon/glasses_storage/episode_1/cam_left_wrist/frame_000435.jpg", "images/agilex_newdragon/glasses_storage/episode_1/cam_right_wrist/frame_000435.jpg", "images/agilex_newdragon/glasses_storage/episode_1/cam_high/frame_000450.jpg", "images/agilex_newdragon/glasses_storage/episode_1/cam_left_wrist/frame_000450.jpg", "images/agilex_newdragon/glasses_storage/episode_1/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.9, "hop_absolute_value": 2.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_16-train-bf_000000-af_000480-plus-0000522", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_high/frame_000538.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_left_wrist/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_right_wrist/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_high/frame_000480.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_left_wrist/frame_000480.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_right_wrist/frame_000480.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 91.4, "hop_absolute_value": 91.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_21-train-bf_000000-af_000540-plus-0000509", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_14/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_14/cam_high/frame_000508.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_left_wrist/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_right_wrist/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_high/frame_000540.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_left_wrist/frame_000540.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_right_wrist/frame_000540.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 78.3, "hop_absolute_value": 78.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_13-train-bf_000030-af_000075-plus-0000682", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_7/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_7/cam_high/frame_001919.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_high/frame_000030.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_left_wrist/frame_000030.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_right_wrist/frame_000030.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_high/frame_000075.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_left_wrist/frame_000075.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.8, "hop_absolute_value": 2.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_14-train-bf_000450-af_000480-plus-0000362", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_18/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_18/cam_high/frame_000414.jpg", "images/agilex_newdragon/glasses_case/episode_14/cam_high/frame_000450.jpg", "images/agilex_newdragon/glasses_case/episode_14/cam_left_wrist/frame_000450.jpg", "images/agilex_newdragon/glasses_case/episode_14/cam_right_wrist/frame_000450.jpg", "images/agilex_newdragon/glasses_case/episode_14/cam_high/frame_000480.jpg", "images/agilex_newdragon/glasses_case/episode_14/cam_left_wrist/frame_000480.jpg", "images/agilex_newdragon/glasses_case/episode_14/cam_right_wrist/frame_000480.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_12-train-bf_000435-af_000450-plus-0000732", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_7/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_7/cam_high/frame_000811.jpg", "images/agilex_newdragon/unplug/episode_12/cam_high/frame_000435.jpg", "images/agilex_newdragon/unplug/episode_12/cam_left_wrist/frame_000435.jpg", "images/agilex_newdragon/unplug/episode_12/cam_right_wrist/frame_000435.jpg", "images/agilex_newdragon/unplug/episode_12/cam_high/frame_000450.jpg", "images/agilex_newdragon/unplug/episode_12/cam_left_wrist/frame_000450.jpg", "images/agilex_newdragon/unplug/episode_12/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_7-train-bf_000615-af_000690-plus-0000358", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_22/cam_high/frame_000651.jpg", "images/agilex_newdragon/unplug/episode_7/cam_high/frame_000615.jpg", "images/agilex_newdragon/unplug/episode_7/cam_left_wrist/frame_000615.jpg", "images/agilex_newdragon/unplug/episode_7/cam_right_wrist/frame_000615.jpg", "images/agilex_newdragon/unplug/episode_7/cam_high/frame_000690.jpg", "images/agilex_newdragon/unplug/episode_7/cam_left_wrist/frame_000690.jpg", "images/agilex_newdragon/unplug/episode_7/cam_right_wrist/frame_000690.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 38.5, "hop_absolute_value": 38.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_0-train-bf_000165-af_000285-plus-0000340", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_22/cam_high/frame_000781.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_0/cam_high/frame_000165.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_0/cam_left_wrist/frame_000165.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_0/cam_right_wrist/frame_000165.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_0/cam_high/frame_000285.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_0/cam_left_wrist/frame_000285.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_0/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 36.4, "hop_absolute_value": 36.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_13-train-bf_000015-af_000330-plus-0000445", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_3/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_3/cam_high/frame_000527.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_high/frame_000015.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_left_wrist/frame_000015.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_right_wrist/frame_000015.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_high/frame_000330.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_left_wrist/frame_000330.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_13/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_15-train-bf_000195-af_000225-plus-0000770", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_14/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_14/cam_high/frame_000902.jpg", "images/agilex_newdragon/unplug/episode_15/cam_high/frame_000195.jpg", "images/agilex_newdragon/unplug/episode_15/cam_left_wrist/frame_000195.jpg", "images/agilex_newdragon/unplug/episode_15/cam_right_wrist/frame_000195.jpg", "images/agilex_newdragon/unplug/episode_15/cam_high/frame_000225.jpg", "images/agilex_newdragon/unplug/episode_15/cam_left_wrist/frame_000225.jpg", "images/agilex_newdragon/unplug/episode_15/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.9, "hop_absolute_value": 2.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_1-train-bf_000780-af_000870-plus-0000387", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_5/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_high/frame_000986.jpg", "images/agilex_newdragon/food_packaging/episode_1/cam_high/frame_000780.jpg", "images/agilex_newdragon/food_packaging/episode_1/cam_left_wrist/frame_000780.jpg", "images/agilex_newdragon/food_packaging/episode_1/cam_right_wrist/frame_000780.jpg", "images/agilex_newdragon/food_packaging/episode_1/cam_high/frame_000870.jpg", "images/agilex_newdragon/food_packaging/episode_1/cam_left_wrist/frame_000870.jpg", "images/agilex_newdragon/food_packaging/episode_1/cam_right_wrist/frame_000870.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 46.2, "hop_absolute_value": 46.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_14-train-bf_000000-af_000090-plus-0000707", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_20/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_20/cam_high/frame_000387.jpg", "images/agilex_newdragon/glasses_case/episode_14/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_14/cam_left_wrist/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_14/cam_right_wrist/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_14/cam_high/frame_000090.jpg", "images/agilex_newdragon/glasses_case/episode_14/cam_left_wrist/frame_000090.jpg", "images/agilex_newdragon/glasses_case/episode_14/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 17.6, "hop_absolute_value": 17.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_20-train-bf_000330-af_000375-plus-0000787", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_6/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_6/cam_high/frame_000468.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_20/cam_high/frame_000330.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_20/cam_left_wrist/frame_000330.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_20/cam_right_wrist/frame_000330.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_20/cam_high/frame_000375.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_20/cam_left_wrist/frame_000375.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_20/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_21-train-bf_000465-af_001635-plus-0000536", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_high/frame_001955.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_high/frame_000465.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_left_wrist/frame_000465.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_right_wrist/frame_000465.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_high/frame_001635.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_left_wrist/frame_001635.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_right_wrist/frame_001635.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 78.8, "hop_absolute_value": 78.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_14-train-bf_000015-af_000030-plus-0000651", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_8/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_8/cam_high/frame_001469.jpg", "images/agilex_newdragon/pour_coffee/episode_14/cam_high/frame_000015.jpg", "images/agilex_newdragon/pour_coffee/episode_14/cam_left_wrist/frame_000015.jpg", "images/agilex_newdragon/pour_coffee/episode_14/cam_right_wrist/frame_000015.jpg", "images/agilex_newdragon/pour_coffee/episode_14/cam_high/frame_000030.jpg", "images/agilex_newdragon/pour_coffee/episode_14/cam_left_wrist/frame_000030.jpg", "images/agilex_newdragon/pour_coffee/episode_14/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.3, "hop_absolute_value": 1.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_22-train-bf_000195-af_000585-plus-0000534", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_8/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_8/cam_high/frame_000490.jpg", "images/agilex_newdragon/unplug/episode_22/cam_high/frame_000195.jpg", "images/agilex_newdragon/unplug/episode_22/cam_left_wrist/frame_000195.jpg", "images/agilex_newdragon/unplug/episode_22/cam_right_wrist/frame_000195.jpg", "images/agilex_newdragon/unplug/episode_22/cam_high/frame_000585.jpg", "images/agilex_newdragon/unplug/episode_22/cam_left_wrist/frame_000585.jpg", "images/agilex_newdragon/unplug/episode_22/cam_right_wrist/frame_000585.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 86.7, "hop_absolute_value": 86.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_20-train-bf_000105-af_000810-plus-0000438", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_20/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_high/frame_001461.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_high/frame_000105.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_left_wrist/frame_000105.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_right_wrist/frame_000105.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_high/frame_000810.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_left_wrist/frame_000810.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_right_wrist/frame_000810.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 52.2, "hop_absolute_value": 52.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_5-train-bf_000870-af_000960-plus-0000542", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_5/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_high/frame_000986.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_high/frame_000870.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_left_wrist/frame_000870.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_right_wrist/frame_000870.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_high/frame_000960.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_left_wrist/frame_000960.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_right_wrist/frame_000960.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_24-train-bf_000150-af_000670-plus-0000584", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_5/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_5/cam_high/frame_000673.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_24/cam_high/frame_000150.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_24/cam_left_wrist/frame_000150.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_24/cam_right_wrist/frame_000150.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_24/cam_high/frame_000670.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_24/cam_left_wrist/frame_000670.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_24/cam_right_wrist/frame_000670.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_6-train-bf_000225-af_000990-plus-0000548", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_high/frame_001698.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_high/frame_000225.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_left_wrist/frame_000225.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_right_wrist/frame_000225.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_high/frame_000990.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_left_wrist/frame_000990.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_right_wrist/frame_000990.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 86.4, "hop_absolute_value": 86.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_13-train-bf_001485-af_001519-plus-0000593", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_high/frame_001519.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_high/frame_001485.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_left_wrist/frame_001485.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_right_wrist/frame_001485.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_high/frame_001519.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_left_wrist/frame_001519.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_right_wrist/frame_001519.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_16-train-bf_000375-af_000405-plus-0000799", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_17/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_17/cam_high/frame_000672.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_high/frame_000375.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_left_wrist/frame_000375.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_right_wrist/frame_000375.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_high/frame_000405.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_left_wrist/frame_000405.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_16/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_5-train-bf_000465-af_000533-plus-0000528", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_2/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_2/cam_high/frame_000480.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_5/cam_high/frame_000465.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_5/cam_left_wrist/frame_000465.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_5/cam_right_wrist/frame_000465.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_5/cam_high/frame_000533.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_5/cam_left_wrist/frame_000533.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_5/cam_right_wrist/frame_000533.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_19-train-bf_000765-af_000915-plus-0000373", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_4/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_4/cam_high/frame_001318.jpg", "images/agilex_newdragon/pour_coffee/episode_19/cam_high/frame_000765.jpg", "images/agilex_newdragon/pour_coffee/episode_19/cam_left_wrist/frame_000765.jpg", "images/agilex_newdragon/pour_coffee/episode_19/cam_right_wrist/frame_000765.jpg", "images/agilex_newdragon/pour_coffee/episode_19/cam_high/frame_000915.jpg", "images/agilex_newdragon/pour_coffee/episode_19/cam_left_wrist/frame_000915.jpg", "images/agilex_newdragon/pour_coffee/episode_19/cam_right_wrist/frame_000915.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 41.7, "hop_absolute_value": 41.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_14-train-bf_000015-af_000150-plus-0000811", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_14/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_14/cam_high/frame_000508.jpg", "images/agilex_newdragon/plug_insertion/episode_14/cam_high/frame_000015.jpg", "images/agilex_newdragon/plug_insertion/episode_14/cam_left_wrist/frame_000015.jpg", "images/agilex_newdragon/plug_insertion/episode_14/cam_right_wrist/frame_000015.jpg", "images/agilex_newdragon/plug_insertion/episode_14/cam_high/frame_000150.jpg", "images/agilex_newdragon/plug_insertion/episode_14/cam_left_wrist/frame_000150.jpg", "images/agilex_newdragon/plug_insertion/episode_14/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 28.1, "hop_absolute_value": 28.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_23-train-bf_000510-af_000525-plus-0000669", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_high/frame_001673.jpg", "images/agilex_newdragon/fold_the_clothes/episode_23/cam_high/frame_000510.jpg", "images/agilex_newdragon/fold_the_clothes/episode_23/cam_left_wrist/frame_000510.jpg", "images/agilex_newdragon/fold_the_clothes/episode_23/cam_right_wrist/frame_000510.jpg", "images/agilex_newdragon/fold_the_clothes/episode_23/cam_high/frame_000525.jpg", "images/agilex_newdragon/fold_the_clothes/episode_23/cam_left_wrist/frame_000525.jpg", "images/agilex_newdragon/fold_the_clothes/episode_23/cam_right_wrist/frame_000525.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.2, "hop_absolute_value": 1.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_17-train-bf_000765-af_001125-plus-0000405", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_17/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_high/frame_001586.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_high/frame_000765.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_left_wrist/frame_000765.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_right_wrist/frame_000765.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_high/frame_001125.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_left_wrist/frame_001125.jpg", "images/agilex_newdragon/food_packaging/episode_17/cam_right_wrist/frame_001125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 44.4, "hop_absolute_value": 44.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_18-train-bf_000255-af_001035-plus-0000435", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_3/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_3/cam_high/frame_001517.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_high/frame_000255.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_left_wrist/frame_000255.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_right_wrist/frame_000255.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_high/frame_001035.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_left_wrist/frame_001035.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_right_wrist/frame_001035.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 52.0, "hop_absolute_value": 52.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_15-train-bf_001245-af_001260-plus-0000669", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_6/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_6/cam_high/frame_002224.jpg", "images/agilex_newdragon/fold_the_clothes/episode_15/cam_high/frame_001245.jpg", "images/agilex_newdragon/fold_the_clothes/episode_15/cam_left_wrist/frame_001245.jpg", "images/agilex_newdragon/fold_the_clothes/episode_15/cam_right_wrist/frame_001245.jpg", "images/agilex_newdragon/fold_the_clothes/episode_15/cam_high/frame_001260.jpg", "images/agilex_newdragon/fold_the_clothes/episode_15/cam_left_wrist/frame_001260.jpg", "images/agilex_newdragon/fold_the_clothes/episode_15/cam_right_wrist/frame_001260.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.8, "hop_absolute_value": 2.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_13-train-bf_000615-af_000630-plus-0000317", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_5/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_high/frame_001409.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_high/frame_000615.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_left_wrist/frame_000615.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_right_wrist/frame_000615.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_high/frame_000630.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_left_wrist/frame_000630.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_right_wrist/frame_000630.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.7, "hop_absolute_value": 1.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_0-train-bf_000195-af_000285-plus-0000326", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_16/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_16/cam_high/frame_000468.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_0/cam_high/frame_000195.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_0/cam_left_wrist/frame_000195.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_0/cam_right_wrist/frame_000195.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_0/cam_high/frame_000285.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_0/cam_left_wrist/frame_000285.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_0/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 21.4, "hop_absolute_value": 21.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_22-train-bf_000555-af_000570-plus-0000830", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_12/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_12/cam_high/frame_000445.jpg", "images/agilex_newdragon/plug_insertion/episode_22/cam_high/frame_000555.jpg", "images/agilex_newdragon/plug_insertion/episode_22/cam_left_wrist/frame_000555.jpg", "images/agilex_newdragon/plug_insertion/episode_22/cam_right_wrist/frame_000555.jpg", "images/agilex_newdragon/plug_insertion/episode_22/cam_high/frame_000570.jpg", "images/agilex_newdragon/plug_insertion/episode_22/cam_left_wrist/frame_000570.jpg", "images/agilex_newdragon/plug_insertion/episode_22/cam_right_wrist/frame_000570.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_23-train-bf_000990-af_001005-plus-0000822", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_23/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_23/cam_high/frame_001634.jpg", "images/agilex_newdragon/food_packaging/episode_23/cam_high/frame_000990.jpg", "images/agilex_newdragon/food_packaging/episode_23/cam_left_wrist/frame_000990.jpg", "images/agilex_newdragon/food_packaging/episode_23/cam_right_wrist/frame_000990.jpg", "images/agilex_newdragon/food_packaging/episode_23/cam_high/frame_001005.jpg", "images/agilex_newdragon/food_packaging/episode_23/cam_left_wrist/frame_001005.jpg", "images/agilex_newdragon/food_packaging/episode_23/cam_right_wrist/frame_001005.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.4, "hop_absolute_value": 2.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_1-train-bf_000285-af_000315-plus-0000880", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_12/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_12/cam_high/frame_001089.jpg", "images/agilex_newdragon/glasses_storage/episode_1/cam_high/frame_000285.jpg", "images/agilex_newdragon/glasses_storage/episode_1/cam_left_wrist/frame_000285.jpg", "images/agilex_newdragon/glasses_storage/episode_1/cam_right_wrist/frame_000285.jpg", "images/agilex_newdragon/glasses_storage/episode_1/cam_high/frame_000315.jpg", "images/agilex_newdragon/glasses_storage/episode_1/cam_left_wrist/frame_000315.jpg", "images/agilex_newdragon/glasses_storage/episode_1/cam_right_wrist/frame_000315.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.4, "hop_absolute_value": 4.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_4-train-bf_000750-af_000765-plus-0000779", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_4/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_4/cam_high/frame_001335.jpg", "images/agilex_newdragon/pour_water/episode_4/cam_high/frame_000750.jpg", "images/agilex_newdragon/pour_water/episode_4/cam_left_wrist/frame_000750.jpg", "images/agilex_newdragon/pour_water/episode_4/cam_right_wrist/frame_000750.jpg", "images/agilex_newdragon/pour_water/episode_4/cam_high/frame_000765.jpg", "images/agilex_newdragon/pour_water/episode_4/cam_left_wrist/frame_000765.jpg", "images/agilex_newdragon/pour_water/episode_4/cam_right_wrist/frame_000765.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.6, "hop_absolute_value": 2.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_19-train-bf_000885-af_000915-plus-0000835", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_24/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_24/cam_high/frame_001828.jpg", "images/agilex_newdragon/fold_the_clothes/episode_19/cam_high/frame_000885.jpg", "images/agilex_newdragon/fold_the_clothes/episode_19/cam_left_wrist/frame_000885.jpg", "images/agilex_newdragon/fold_the_clothes/episode_19/cam_right_wrist/frame_000885.jpg", "images/agilex_newdragon/fold_the_clothes/episode_19/cam_high/frame_000915.jpg", "images/agilex_newdragon/fold_the_clothes/episode_19/cam_left_wrist/frame_000915.jpg", "images/agilex_newdragon/fold_the_clothes/episode_19/cam_right_wrist/frame_000915.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.8, "hop_absolute_value": 2.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_23-train-bf_000270-af_000285-plus-0000278", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_4/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_high/frame_000413.jpg", "images/agilex_newdragon/glasses_case/episode_23/cam_high/frame_000270.jpg", "images/agilex_newdragon/glasses_case/episode_23/cam_left_wrist/frame_000270.jpg", "images/agilex_newdragon/glasses_case/episode_23/cam_right_wrist/frame_000270.jpg", "images/agilex_newdragon/glasses_case/episode_23/cam_high/frame_000285.jpg", "images/agilex_newdragon/glasses_case/episode_23/cam_left_wrist/frame_000285.jpg", "images/agilex_newdragon/glasses_case/episode_23/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_16-train-bf_000540-af_000585-plus-0000948", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_18/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_18/cam_high/frame_001159.jpg", "images/agilex_newdragon/unplug/episode_16/cam_high/frame_000540.jpg", "images/agilex_newdragon/unplug/episode_16/cam_left_wrist/frame_000540.jpg", "images/agilex_newdragon/unplug/episode_16/cam_right_wrist/frame_000540.jpg", "images/agilex_newdragon/unplug/episode_16/cam_high/frame_000585.jpg", "images/agilex_newdragon/unplug/episode_16/cam_left_wrist/frame_000585.jpg", "images/agilex_newdragon/unplug/episode_16/cam_right_wrist/frame_000585.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_11-train-bf_000225-af_000240-plus-0000859", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_high/frame_001673.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_000225.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_left_wrist/frame_000225.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_right_wrist/frame_000225.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_000240.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_left_wrist/frame_000240.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 0.5, "hop_absolute_value": 0.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_13-train-bf_001125-af_001140-plus-0000765", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_4/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_4/cam_high/frame_001929.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_high/frame_001125.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_left_wrist/frame_001125.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_right_wrist/frame_001125.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_high/frame_001140.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_left_wrist/frame_001140.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_right_wrist/frame_001140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.8, "hop_absolute_value": 2.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_8-train-bf_000090-af_000105-plus-0000803", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_12/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_12/cam_high/frame_001711.jpg", "images/agilex_newdragon/clean_the_desk/episode_8/cam_high/frame_000090.jpg", "images/agilex_newdragon/clean_the_desk/episode_8/cam_left_wrist/frame_000090.jpg", "images/agilex_newdragon/clean_the_desk/episode_8/cam_right_wrist/frame_000090.jpg", "images/agilex_newdragon/clean_the_desk/episode_8/cam_high/frame_000105.jpg", "images/agilex_newdragon/clean_the_desk/episode_8/cam_left_wrist/frame_000105.jpg", "images/agilex_newdragon/clean_the_desk/episode_8/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.0, "hop_absolute_value": 1.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_21-train-bf_000750-af_000773-plus-0000566", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_22/cam_high/frame_000948.jpg", "images/agilex_newdragon/glasses_storage/episode_21/cam_high/frame_000750.jpg", "images/agilex_newdragon/glasses_storage/episode_21/cam_left_wrist/frame_000750.jpg", "images/agilex_newdragon/glasses_storage/episode_21/cam_right_wrist/frame_000750.jpg", "images/agilex_newdragon/glasses_storage/episode_21/cam_high/frame_000773.jpg", "images/agilex_newdragon/glasses_storage/episode_21/cam_left_wrist/frame_000773.jpg", "images/agilex_newdragon/glasses_storage/episode_21/cam_right_wrist/frame_000773.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_15-train-bf_000660-af_000675-plus-0000884", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_15/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_high/frame_001642.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_high/frame_000660.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_left_wrist/frame_000660.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_right_wrist/frame_000660.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_high/frame_000675.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_left_wrist/frame_000675.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_right_wrist/frame_000675.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.5, "hop_absolute_value": 1.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_6-train-bf_000240-af_000270-plus-0000880", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_19/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_19/cam_high/frame_001165.jpg", "images/agilex_newdragon/unplug/episode_6/cam_high/frame_000240.jpg", "images/agilex_newdragon/unplug/episode_6/cam_left_wrist/frame_000240.jpg", "images/agilex_newdragon/unplug/episode_6/cam_right_wrist/frame_000240.jpg", "images/agilex_newdragon/unplug/episode_6/cam_high/frame_000270.jpg", "images/agilex_newdragon/unplug/episode_6/cam_left_wrist/frame_000270.jpg", "images/agilex_newdragon/unplug/episode_6/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.1, "hop_absolute_value": 5.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_4-train-bf_001005-af_001020-plus-0000789", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_4/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_4/cam_high/frame_001929.jpg", "images/agilex_newdragon/fold_the_clothes/episode_4/cam_high/frame_001005.jpg", "images/agilex_newdragon/fold_the_clothes/episode_4/cam_left_wrist/frame_001005.jpg", "images/agilex_newdragon/fold_the_clothes/episode_4/cam_right_wrist/frame_001005.jpg", "images/agilex_newdragon/fold_the_clothes/episode_4/cam_high/frame_001020.jpg", "images/agilex_newdragon/fold_the_clothes/episode_4/cam_left_wrist/frame_001020.jpg", "images/agilex_newdragon/fold_the_clothes/episode_4/cam_right_wrist/frame_001020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.6, "hop_absolute_value": 1.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_17-train-bf_001755-af_001785-plus-0000337", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_17/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_17/cam_high/frame_001963.jpg", "images/agilex_newdragon/fold_the_clothes/episode_17/cam_high/frame_001755.jpg", "images/agilex_newdragon/fold_the_clothes/episode_17/cam_left_wrist/frame_001755.jpg", "images/agilex_newdragon/fold_the_clothes/episode_17/cam_right_wrist/frame_001755.jpg", "images/agilex_newdragon/fold_the_clothes/episode_17/cam_high/frame_001785.jpg", "images/agilex_newdragon/fold_the_clothes/episode_17/cam_left_wrist/frame_001785.jpg", "images/agilex_newdragon/fold_the_clothes/episode_17/cam_right_wrist/frame_001785.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 15.4, "hop_absolute_value": 15.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_9-train-bf_000390-af_000405-plus-0000301", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_11/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_11/cam_high/frame_000595.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_high/frame_000390.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_left_wrist/frame_000390.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_right_wrist/frame_000390.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_high/frame_000405.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_left_wrist/frame_000405.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_11-train-bf_001575-af_001590-plus-0000949", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_5/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_5/cam_high/frame_001845.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_001575.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_left_wrist/frame_001575.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_right_wrist/frame_001575.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_001590.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_left_wrist/frame_001590.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_right_wrist/frame_001590.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 0.9, "hop_absolute_value": 0.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_7-train-bf_000075-af_000000-minus-0000002", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_12/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_12/cam_high/frame_000541.jpg", "images/agilex_newdragon/glasses_case/episode_7/cam_high/frame_000075.jpg", "images/agilex_newdragon/glasses_case/episode_7/cam_left_wrist/frame_000075.jpg", "images/agilex_newdragon/glasses_case/episode_7/cam_right_wrist/frame_000075.jpg", "images/agilex_newdragon/glasses_case/episode_7/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_7/cam_left_wrist/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_7/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_1-train-bf_000660-af_000330-minus-0000207", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_high/frame_001955.jpg", "images/agilex_newdragon/fold_the_clothes/episode_1/cam_high/frame_000660.jpg", "images/agilex_newdragon/fold_the_clothes/episode_1/cam_left_wrist/frame_000660.jpg", "images/agilex_newdragon/fold_the_clothes/episode_1/cam_right_wrist/frame_000660.jpg", "images/agilex_newdragon/fold_the_clothes/episode_1/cam_high/frame_000330.jpg", "images/agilex_newdragon/fold_the_clothes/episode_1/cam_left_wrist/frame_000330.jpg", "images/agilex_newdragon/fold_the_clothes/episode_1/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_8-train-bf_001260-af_000525-minus-0000200", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_14/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_14/cam_high/frame_001200.jpg", "images/agilex_newdragon/pour_coffee/episode_8/cam_high/frame_001260.jpg", "images/agilex_newdragon/pour_coffee/episode_8/cam_left_wrist/frame_001260.jpg", "images/agilex_newdragon/pour_coffee/episode_8/cam_right_wrist/frame_001260.jpg", "images/agilex_newdragon/pour_coffee/episode_8/cam_high/frame_000525.jpg", "images/agilex_newdragon/pour_coffee/episode_8/cam_left_wrist/frame_000525.jpg", "images/agilex_newdragon/pour_coffee/episode_8/cam_right_wrist/frame_000525.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -58.3, "hop_absolute_value": 58.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_20-train-bf_000525-af_000135-minus-0000125", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_20/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_20/cam_high/frame_000798.jpg", "images/agilex_newdragon/glasses_storage/episode_20/cam_high/frame_000525.jpg", "images/agilex_newdragon/glasses_storage/episode_20/cam_left_wrist/frame_000525.jpg", "images/agilex_newdragon/glasses_storage/episode_20/cam_right_wrist/frame_000525.jpg", "images/agilex_newdragon/glasses_storage/episode_20/cam_high/frame_000135.jpg", "images/agilex_newdragon/glasses_storage/episode_20/cam_left_wrist/frame_000135.jpg", "images/agilex_newdragon/glasses_storage/episode_20/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -74.3, "hop_absolute_value": 74.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_8-train-bf_000720-af_000630-minus-0000284", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_2/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_2/cam_high/frame_000919.jpg", "images/agilex_newdragon/food_packaging/episode_8/cam_high/frame_000720.jpg", "images/agilex_newdragon/food_packaging/episode_8/cam_left_wrist/frame_000720.jpg", "images/agilex_newdragon/food_packaging/episode_8/cam_right_wrist/frame_000720.jpg", "images/agilex_newdragon/food_packaging/episode_8/cam_high/frame_000630.jpg", "images/agilex_newdragon/food_packaging/episode_8/cam_left_wrist/frame_000630.jpg", "images/agilex_newdragon/food_packaging/episode_8/cam_right_wrist/frame_000630.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_7-train-bf_000690-af_000630-minus-0000750", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_9/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_9/cam_high/frame_001143.jpg", "images/agilex_newdragon/food_packaging/episode_7/cam_high/frame_000690.jpg", "images/agilex_newdragon/food_packaging/episode_7/cam_left_wrist/frame_000690.jpg", "images/agilex_newdragon/food_packaging/episode_7/cam_right_wrist/frame_000690.jpg", "images/agilex_newdragon/food_packaging/episode_7/cam_high/frame_000630.jpg", "images/agilex_newdragon/food_packaging/episode_7/cam_left_wrist/frame_000630.jpg", "images/agilex_newdragon/food_packaging/episode_7/cam_right_wrist/frame_000630.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -8.7, "hop_absolute_value": 8.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_5-train-bf_000915-af_000090-minus-0000058", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_14/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_14/cam_high/frame_001771.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_high/frame_000915.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_left_wrist/frame_000915.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_right_wrist/frame_000915.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_high/frame_000090.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_left_wrist/frame_000090.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -90.2, "hop_absolute_value": 90.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_24-train-bf_000566-af_000270-minus-0000210", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_13/cam_high/frame_001676.jpg", "images/agilex_newdragon/pour_water/episode_24/cam_high/frame_000566.jpg", "images/agilex_newdragon/pour_water/episode_24/cam_left_wrist/frame_000566.jpg", "images/agilex_newdragon/pour_water/episode_24/cam_right_wrist/frame_000566.jpg", "images/agilex_newdragon/pour_water/episode_24/cam_high/frame_000270.jpg", "images/agilex_newdragon/pour_water/episode_24/cam_left_wrist/frame_000270.jpg", "images/agilex_newdragon/pour_water/episode_24/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -51.4, "hop_absolute_value": 51.4, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_4-train-bf_001020-af_000930-minus-0000302", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_12/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_12/cam_high/frame_001505.jpg", "images/agilex_newdragon/pour_coffee/episode_4/cam_high/frame_001020.jpg", "images/agilex_newdragon/pour_coffee/episode_4/cam_left_wrist/frame_001020.jpg", "images/agilex_newdragon/pour_coffee/episode_4/cam_right_wrist/frame_001020.jpg", "images/agilex_newdragon/pour_coffee/episode_4/cam_high/frame_000930.jpg", "images/agilex_newdragon/pour_coffee/episode_4/cam_left_wrist/frame_000930.jpg", "images/agilex_newdragon/pour_coffee/episode_4/cam_right_wrist/frame_000930.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -8.8, "hop_absolute_value": 8.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_23-train-bf_000600-af_000345-minus-0000226", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_10/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_10/cam_high/frame_001439.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_high/frame_000600.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_left_wrist/frame_000600.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_right_wrist/frame_000600.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_high/frame_000345.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_left_wrist/frame_000345.jpg", "images/agilex_newdragon/pour_coffee/episode_23/cam_right_wrist/frame_000345.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -42.5, "hop_absolute_value": 42.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_1-train-bf_000315-af_000210-minus-0000207", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_13/cam_high/frame_000622.jpg", "images/agilex_newdragon/glasses_case/episode_1/cam_high/frame_000315.jpg", "images/agilex_newdragon/glasses_case/episode_1/cam_left_wrist/frame_000315.jpg", "images/agilex_newdragon/glasses_case/episode_1/cam_right_wrist/frame_000315.jpg", "images/agilex_newdragon/glasses_case/episode_1/cam_high/frame_000210.jpg", "images/agilex_newdragon/glasses_case/episode_1/cam_left_wrist/frame_000210.jpg", "images/agilex_newdragon/glasses_case/episode_1/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_2-train-bf_000045-af_000030-minus-0000648", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_19/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_19/cam_high/frame_000477.jpg", "images/agilex_newdragon/glasses_case/episode_2/cam_high/frame_000045.jpg", "images/agilex_newdragon/glasses_case/episode_2/cam_left_wrist/frame_000045.jpg", "images/agilex_newdragon/glasses_case/episode_2/cam_right_wrist/frame_000045.jpg", "images/agilex_newdragon/glasses_case/episode_2/cam_high/frame_000030.jpg", "images/agilex_newdragon/glasses_case/episode_2/cam_left_wrist/frame_000030.jpg", "images/agilex_newdragon/glasses_case/episode_2/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_5-train-bf_000915-af_000900-minus-0000307", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_5/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_high/frame_001409.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_high/frame_000915.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_left_wrist/frame_000915.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_right_wrist/frame_000915.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_high/frame_000900.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_left_wrist/frame_000900.jpg", "images/agilex_newdragon/clean_the_desk/episode_5/cam_right_wrist/frame_000900.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.6, "hop_absolute_value": 1.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_1-train-bf_000735-af_000720-minus-0000657", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_3/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_3/cam_high/frame_001783.jpg", "images/agilex_newdragon/fold_the_clothes/episode_1/cam_high/frame_000735.jpg", "images/agilex_newdragon/fold_the_clothes/episode_1/cam_left_wrist/frame_000735.jpg", "images/agilex_newdragon/fold_the_clothes/episode_1/cam_right_wrist/frame_000735.jpg", "images/agilex_newdragon/fold_the_clothes/episode_1/cam_high/frame_000720.jpg", "images/agilex_newdragon/fold_the_clothes/episode_1/cam_left_wrist/frame_000720.jpg", "images/agilex_newdragon/fold_the_clothes/episode_1/cam_right_wrist/frame_000720.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.0, "hop_absolute_value": 2.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_15-train-bf_000285-af_000015-minus-0000024", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_18/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_18/cam_high/frame_000703.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_15/cam_high/frame_000285.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_15/cam_left_wrist/frame_000285.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_15/cam_right_wrist/frame_000285.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_15/cam_high/frame_000015.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_15/cam_left_wrist/frame_000015.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_15/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -94.7, "hop_absolute_value": 94.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_10-train-bf_002220-af_000420-minus-0000108", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_18/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_18/cam_high/frame_001625.jpg", "images/agilex_newdragon/fold_the_clothes/episode_10/cam_high/frame_002220.jpg", "images/agilex_newdragon/fold_the_clothes/episode_10/cam_left_wrist/frame_002220.jpg", "images/agilex_newdragon/fold_the_clothes/episode_10/cam_right_wrist/frame_002220.jpg", "images/agilex_newdragon/fold_the_clothes/episode_10/cam_high/frame_000420.jpg", "images/agilex_newdragon/fold_the_clothes/episode_10/cam_left_wrist/frame_000420.jpg", "images/agilex_newdragon/fold_the_clothes/episode_10/cam_right_wrist/frame_000420.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -81.1, "hop_absolute_value": 81.1, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_6-train-bf_000150-af_000135-minus-0000570", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_12/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_12/cam_high/frame_000888.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_6/cam_high/frame_000150.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_6/cam_left_wrist/frame_000150.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_6/cam_right_wrist/frame_000150.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_6/cam_high/frame_000135.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_6/cam_left_wrist/frame_000135.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_6/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_16-train-bf_000345-af_000105-minus-0000121", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_5/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_5/cam_high/frame_000494.jpg", "images/agilex_newdragon/plug_insertion/episode_16/cam_high/frame_000345.jpg", "images/agilex_newdragon/plug_insertion/episode_16/cam_left_wrist/frame_000345.jpg", "images/agilex_newdragon/plug_insertion/episode_16/cam_right_wrist/frame_000345.jpg", "images/agilex_newdragon/plug_insertion/episode_16/cam_high/frame_000105.jpg", "images/agilex_newdragon/plug_insertion/episode_16/cam_left_wrist/frame_000105.jpg", "images/agilex_newdragon/plug_insertion/episode_16/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -69.6, "hop_absolute_value": 69.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_4-train-bf_000285-af_000090-minus-0000099", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_9/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_9/cam_high/frame_000563.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_high/frame_000285.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_left_wrist/frame_000285.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_right_wrist/frame_000285.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_high/frame_000090.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_left_wrist/frame_000090.jpg", "images/agilex_newdragon/glasses_case/episode_4/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -68.4, "hop_absolute_value": 68.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_2-train-bf_000360-af_000015-minus-0000005", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_24/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_24/cam_high/frame_001828.jpg", "images/agilex_newdragon/fold_the_clothes/episode_2/cam_high/frame_000360.jpg", "images/agilex_newdragon/fold_the_clothes/episode_2/cam_left_wrist/frame_000360.jpg", "images/agilex_newdragon/fold_the_clothes/episode_2/cam_right_wrist/frame_000360.jpg", "images/agilex_newdragon/fold_the_clothes/episode_2/cam_high/frame_000015.jpg", "images/agilex_newdragon/fold_the_clothes/episode_2/cam_left_wrist/frame_000015.jpg", "images/agilex_newdragon/fold_the_clothes/episode_2/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -95.8, "hop_absolute_value": 95.8, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_21-train-bf_001125-af_000195-minus-0000108", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_21/cam_high/frame_001196.jpg", "images/agilex_newdragon/food_packaging/episode_21/cam_high/frame_001125.jpg", "images/agilex_newdragon/food_packaging/episode_21/cam_left_wrist/frame_001125.jpg", "images/agilex_newdragon/food_packaging/episode_21/cam_right_wrist/frame_001125.jpg", "images/agilex_newdragon/food_packaging/episode_21/cam_high/frame_000195.jpg", "images/agilex_newdragon/food_packaging/episode_21/cam_left_wrist/frame_000195.jpg", "images/agilex_newdragon/food_packaging/episode_21/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -82.7, "hop_absolute_value": 82.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_20-train-bf_001200-af_000180-minus-0000057", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_15/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_high/frame_001642.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_high/frame_001200.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_left_wrist/frame_001200.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_right_wrist/frame_001200.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_high/frame_000180.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_left_wrist/frame_000180.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -85.0, "hop_absolute_value": 85.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_17-train-bf_000465-af_000255-minus-0000207", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_0/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_0/cam_high/frame_000827.jpg", "images/agilex_newdragon/glasses_storage/episode_17/cam_high/frame_000465.jpg", "images/agilex_newdragon/glasses_storage/episode_17/cam_left_wrist/frame_000465.jpg", "images/agilex_newdragon/glasses_storage/episode_17/cam_right_wrist/frame_000465.jpg", "images/agilex_newdragon/glasses_storage/episode_17/cam_high/frame_000255.jpg", "images/agilex_newdragon/glasses_storage/episode_17/cam_left_wrist/frame_000255.jpg", "images/agilex_newdragon/glasses_storage/episode_17/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -45.2, "hop_absolute_value": 45.2, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_13-train-bf_000330-af_000270-minus-0000780", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_20/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_20/cam_high/frame_000626.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_high/frame_000330.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_left_wrist/frame_000330.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_right_wrist/frame_000330.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_high/frame_000270.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_left_wrist/frame_000270.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -18.2, "hop_absolute_value": 18.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_9-train-bf_000510-af_000495-minus-0000751", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_9/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_9/cam_high/frame_001143.jpg", "images/agilex_newdragon/food_packaging/episode_9/cam_high/frame_000510.jpg", "images/agilex_newdragon/food_packaging/episode_9/cam_left_wrist/frame_000510.jpg", "images/agilex_newdragon/food_packaging/episode_9/cam_right_wrist/frame_000510.jpg", "images/agilex_newdragon/food_packaging/episode_9/cam_high/frame_000495.jpg", "images/agilex_newdragon/food_packaging/episode_9/cam_left_wrist/frame_000495.jpg", "images/agilex_newdragon/food_packaging/episode_9/cam_right_wrist/frame_000495.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.9, "hop_absolute_value": 2.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_8-train-bf_000870-af_000855-minus-0000674", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_19/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_19/cam_high/frame_001961.jpg", "images/agilex_newdragon/fold_the_clothes/episode_8/cam_high/frame_000870.jpg", "images/agilex_newdragon/fold_the_clothes/episode_8/cam_left_wrist/frame_000870.jpg", "images/agilex_newdragon/fold_the_clothes/episode_8/cam_right_wrist/frame_000870.jpg", "images/agilex_newdragon/fold_the_clothes/episode_8/cam_high/frame_000855.jpg", "images/agilex_newdragon/fold_the_clothes/episode_8/cam_left_wrist/frame_000855.jpg", "images/agilex_newdragon/fold_the_clothes/episode_8/cam_right_wrist/frame_000855.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.7, "hop_absolute_value": 1.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_18-train-bf_000360-af_000015-minus-0000015", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_23/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_23/cam_high/frame_001634.jpg", "images/agilex_newdragon/food_packaging/episode_18/cam_high/frame_000360.jpg", "images/agilex_newdragon/food_packaging/episode_18/cam_left_wrist/frame_000360.jpg", "images/agilex_newdragon/food_packaging/episode_18/cam_right_wrist/frame_000360.jpg", "images/agilex_newdragon/food_packaging/episode_18/cam_high/frame_000015.jpg", "images/agilex_newdragon/food_packaging/episode_18/cam_left_wrist/frame_000015.jpg", "images/agilex_newdragon/food_packaging/episode_18/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -95.8, "hop_absolute_value": 95.8, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_15-train-bf_000945-af_000840-minus-0000880", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_3/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_3/cam_high/frame_001245.jpg", "images/agilex_newdragon/pour_water/episode_15/cam_high/frame_000945.jpg", "images/agilex_newdragon/pour_water/episode_15/cam_left_wrist/frame_000945.jpg", "images/agilex_newdragon/pour_water/episode_15/cam_right_wrist/frame_000945.jpg", "images/agilex_newdragon/pour_water/episode_15/cam_high/frame_000840.jpg", "images/agilex_newdragon/pour_water/episode_15/cam_left_wrist/frame_000840.jpg", "images/agilex_newdragon/pour_water/episode_15/cam_right_wrist/frame_000840.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_1-train-bf_001065-af_001050-minus-0000306", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_20/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_20/cam_high/frame_001156.jpg", "images/agilex_newdragon/pour_water/episode_1/cam_high/frame_001065.jpg", "images/agilex_newdragon/pour_water/episode_1/cam_left_wrist/frame_001065.jpg", "images/agilex_newdragon/pour_water/episode_1/cam_right_wrist/frame_001065.jpg", "images/agilex_newdragon/pour_water/episode_1/cam_high/frame_001050.jpg", "images/agilex_newdragon/pour_water/episode_1/cam_left_wrist/frame_001050.jpg", "images/agilex_newdragon/pour_water/episode_1/cam_right_wrist/frame_001050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.4, "hop_absolute_value": 1.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_21-train-bf_000135-af_000120-minus-0000649", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_high/frame_000697.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_high/frame_000135.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_left_wrist/frame_000135.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_right_wrist/frame_000135.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_high/frame_000120.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_left_wrist/frame_000120.jpg", "images/agilex_newdragon/plug_insertion/episode_21/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_3-train-bf_000195-af_000075-minus-0000127", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_20/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_20/cam_high/frame_000508.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_high/frame_000195.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_left_wrist/frame_000195.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_right_wrist/frame_000195.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_high/frame_000075.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_left_wrist/frame_000075.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -61.5, "hop_absolute_value": 61.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_3-train-bf_000300-af_000090-minus-0000104", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_3/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_high/frame_000472.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_high/frame_000300.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_left_wrist/frame_000300.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_right_wrist/frame_000300.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_high/frame_000090.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_left_wrist/frame_000090.jpg", "images/agilex_newdragon/plug_insertion/episode_3/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_7-train-bf_000375-af_000165-minus-0000883", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_22/cam_high/frame_000781.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_7/cam_high/frame_000375.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_7/cam_left_wrist/frame_000375.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_7/cam_right_wrist/frame_000375.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_7/cam_high/frame_000165.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_7/cam_left_wrist/frame_000165.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_7/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -56.0, "hop_absolute_value": 56.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_6-train-bf_000750-af_000120-minus-0000103", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_0/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_0/cam_high/frame_000987.jpg", "images/agilex_newdragon/clean_the_desk/episode_6/cam_high/frame_000750.jpg", "images/agilex_newdragon/clean_the_desk/episode_6/cam_left_wrist/frame_000750.jpg", "images/agilex_newdragon/clean_the_desk/episode_6/cam_right_wrist/frame_000750.jpg", "images/agilex_newdragon/clean_the_desk/episode_6/cam_high/frame_000120.jpg", "images/agilex_newdragon/clean_the_desk/episode_6/cam_left_wrist/frame_000120.jpg", "images/agilex_newdragon/clean_the_desk/episode_6/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -84.0, "hop_absolute_value": 84.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_0-train-bf_000315-af_000225-minus-0000901", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_18/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_18/cam_high/frame_001159.jpg", "images/agilex_newdragon/unplug/episode_0/cam_high/frame_000315.jpg", "images/agilex_newdragon/unplug/episode_0/cam_left_wrist/frame_000315.jpg", "images/agilex_newdragon/unplug/episode_0/cam_right_wrist/frame_000315.jpg", "images/agilex_newdragon/unplug/episode_0/cam_high/frame_000225.jpg", "images/agilex_newdragon/unplug/episode_0/cam_left_wrist/frame_000225.jpg", "images/agilex_newdragon/unplug/episode_0/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -28.6, "hop_absolute_value": 28.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_19-train-bf_001245-af_001230-minus-0000691", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_10/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_10/cam_high/frame_001505.jpg", "images/agilex_newdragon/clean_the_desk/episode_19/cam_high/frame_001245.jpg", "images/agilex_newdragon/clean_the_desk/episode_19/cam_left_wrist/frame_001245.jpg", "images/agilex_newdragon/clean_the_desk/episode_19/cam_right_wrist/frame_001245.jpg", "images/agilex_newdragon/clean_the_desk/episode_19/cam_high/frame_001230.jpg", "images/agilex_newdragon/clean_the_desk/episode_19/cam_left_wrist/frame_001230.jpg", "images/agilex_newdragon/clean_the_desk/episode_19/cam_right_wrist/frame_001230.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.2, "hop_absolute_value": 1.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_13-train-bf_000405-af_000225-minus-0000210", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_0/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_high/frame_002049.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_high/frame_000405.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_left_wrist/frame_000405.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_right_wrist/frame_000405.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_high/frame_000225.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_left_wrist/frame_000225.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -44.4, "hop_absolute_value": 44.4, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_3-train-bf_000555-af_000495-minus-0000860", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_1/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_1/cam_high/frame_000984.jpg", "images/agilex_newdragon/food_packaging/episode_3/cam_high/frame_000555.jpg", "images/agilex_newdragon/food_packaging/episode_3/cam_left_wrist/frame_000555.jpg", "images/agilex_newdragon/food_packaging/episode_3/cam_right_wrist/frame_000555.jpg", "images/agilex_newdragon/food_packaging/episode_3/cam_high/frame_000495.jpg", "images/agilex_newdragon/food_packaging/episode_3/cam_left_wrist/frame_000495.jpg", "images/agilex_newdragon/food_packaging/episode_3/cam_right_wrist/frame_000495.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.8, "hop_absolute_value": 10.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_18-train-bf_000435-af_000165-minus-0000142", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_18/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_high/frame_000476.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_high/frame_000435.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_left_wrist/frame_000435.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_right_wrist/frame_000435.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_high/frame_000165.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_left_wrist/frame_000165.jpg", "images/agilex_newdragon/plug_insertion/episode_18/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -62.1, "hop_absolute_value": 62.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_16-train-bf_000795-af_000450-minus-0000235", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_22/cam_high/frame_000948.jpg", "images/agilex_newdragon/glasses_storage/episode_16/cam_high/frame_000795.jpg", "images/agilex_newdragon/glasses_storage/episode_16/cam_left_wrist/frame_000795.jpg", "images/agilex_newdragon/glasses_storage/episode_16/cam_right_wrist/frame_000795.jpg", "images/agilex_newdragon/glasses_storage/episode_16/cam_high/frame_000450.jpg", "images/agilex_newdragon/glasses_storage/episode_16/cam_left_wrist/frame_000450.jpg", "images/agilex_newdragon/glasses_storage/episode_16/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -43.4, "hop_absolute_value": 43.4, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_3-train-bf_000210-af_000195-minus-0000647", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_21/cam_high/frame_000773.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_high/frame_000210.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_left_wrist/frame_000210.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_right_wrist/frame_000210.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_high/frame_000195.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_left_wrist/frame_000195.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.1, "hop_absolute_value": 7.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_12-train-bf_000735-af_000465-minus-0000228", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_13/cam_high/frame_001246.jpg", "images/agilex_newdragon/glasses_storage/episode_12/cam_high/frame_000735.jpg", "images/agilex_newdragon/glasses_storage/episode_12/cam_left_wrist/frame_000735.jpg", "images/agilex_newdragon/glasses_storage/episode_12/cam_right_wrist/frame_000735.jpg", "images/agilex_newdragon/glasses_storage/episode_12/cam_high/frame_000465.jpg", "images/agilex_newdragon/glasses_storage/episode_12/cam_left_wrist/frame_000465.jpg", "images/agilex_newdragon/glasses_storage/episode_12/cam_right_wrist/frame_000465.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -36.7, "hop_absolute_value": 36.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_10-train-bf_001215-af_000555-minus-0000190", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_13/cam_high/frame_001673.jpg", "images/agilex_newdragon/fold_the_clothes/episode_10/cam_high/frame_001215.jpg", "images/agilex_newdragon/fold_the_clothes/episode_10/cam_left_wrist/frame_001215.jpg", "images/agilex_newdragon/fold_the_clothes/episode_10/cam_right_wrist/frame_001215.jpg", "images/agilex_newdragon/fold_the_clothes/episode_10/cam_high/frame_000555.jpg", "images/agilex_newdragon/fold_the_clothes/episode_10/cam_left_wrist/frame_000555.jpg", "images/agilex_newdragon/fold_the_clothes/episode_10/cam_right_wrist/frame_000555.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -54.3, "hop_absolute_value": 54.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_19-train-bf_000330-af_000300-minus-0000890", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_5/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_5/cam_high/frame_000448.jpg", "images/agilex_newdragon/unplug/episode_19/cam_high/frame_000330.jpg", "images/agilex_newdragon/unplug/episode_19/cam_left_wrist/frame_000330.jpg", "images/agilex_newdragon/unplug/episode_19/cam_right_wrist/frame_000330.jpg", "images/agilex_newdragon/unplug/episode_19/cam_high/frame_000300.jpg", "images/agilex_newdragon/unplug/episode_19/cam_left_wrist/frame_000300.jpg", "images/agilex_newdragon/unplug/episode_19/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_clothes_episode_22-train-bf_000765-af_000645-minus-0000275", "task": "fold clothes", "image": [ "images/agilex_newdragon/fold_clothes/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_clothes/episode_22/cam_high/frame_000896.jpg", "images/agilex_newdragon/fold_clothes/episode_22/cam_high/frame_000765.jpg", "images/agilex_newdragon/fold_clothes/episode_22/cam_left_wrist/frame_000765.jpg", "images/agilex_newdragon/fold_clothes/episode_22/cam_right_wrist/frame_000765.jpg", "images/agilex_newdragon/fold_clothes/episode_22/cam_high/frame_000645.jpg", "images/agilex_newdragon/fold_clothes/episode_22/cam_left_wrist/frame_000645.jpg", "images/agilex_newdragon/fold_clothes/episode_22/cam_right_wrist/frame_000645.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -15.7, "hop_absolute_value": 15.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_15-train-bf_001500-af_001080-minus-0000284", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_21/cam_high/frame_001196.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_high/frame_001500.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_left_wrist/frame_001500.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_right_wrist/frame_001500.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_high/frame_001080.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_left_wrist/frame_001080.jpg", "images/agilex_newdragon/food_packaging/episode_15/cam_right_wrist/frame_001080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -28.0, "hop_absolute_value": 28.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_5-train-bf_000345-af_000165-minus-0000167", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_13/cam_high/frame_000562.jpg", "images/agilex_newdragon/plug_insertion/episode_5/cam_high/frame_000345.jpg", "images/agilex_newdragon/plug_insertion/episode_5/cam_left_wrist/frame_000345.jpg", "images/agilex_newdragon/plug_insertion/episode_5/cam_right_wrist/frame_000345.jpg", "images/agilex_newdragon/plug_insertion/episode_5/cam_high/frame_000165.jpg", "images/agilex_newdragon/plug_insertion/episode_5/cam_left_wrist/frame_000165.jpg", "images/agilex_newdragon/plug_insertion/episode_5/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -52.2, "hop_absolute_value": 52.2, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_22-train-bf_000135-af_000030-minus-0000077", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_8/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_8/cam_high/frame_000520.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_22/cam_high/frame_000135.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_22/cam_left_wrist/frame_000135.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_22/cam_right_wrist/frame_000135.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_22/cam_high/frame_000030.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_22/cam_left_wrist/frame_000030.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_22/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -77.8, "hop_absolute_value": 77.8, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_12-train-bf_001425-af_000495-minus-0000166", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_12/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_12/cam_high/frame_001488.jpg", "images/agilex_newdragon/food_packaging/episode_12/cam_high/frame_001425.jpg", "images/agilex_newdragon/food_packaging/episode_12/cam_left_wrist/frame_001425.jpg", "images/agilex_newdragon/food_packaging/episode_12/cam_right_wrist/frame_001425.jpg", "images/agilex_newdragon/food_packaging/episode_12/cam_high/frame_000495.jpg", "images/agilex_newdragon/food_packaging/episode_12/cam_left_wrist/frame_000495.jpg", "images/agilex_newdragon/food_packaging/episode_12/cam_right_wrist/frame_000495.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -65.3, "hop_absolute_value": 65.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_6-train-bf_001050-af_001035-minus-0000694", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_6/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_high/frame_001117.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_high/frame_001050.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_left_wrist/frame_001050.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_right_wrist/frame_001050.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_high/frame_001035.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_left_wrist/frame_001035.jpg", "images/agilex_newdragon/food_packaging/episode_6/cam_right_wrist/frame_001035.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.4, "hop_absolute_value": 1.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_3-train-bf_001155-af_001125-minus-0000774", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_4/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_4/cam_high/frame_001929.jpg", "images/agilex_newdragon/fold_the_clothes/episode_3/cam_high/frame_001155.jpg", "images/agilex_newdragon/fold_the_clothes/episode_3/cam_left_wrist/frame_001155.jpg", "images/agilex_newdragon/fold_the_clothes/episode_3/cam_right_wrist/frame_001155.jpg", "images/agilex_newdragon/fold_the_clothes/episode_3/cam_high/frame_001125.jpg", "images/agilex_newdragon/fold_the_clothes/episode_3/cam_left_wrist/frame_001125.jpg", "images/agilex_newdragon/fold_the_clothes/episode_3/cam_right_wrist/frame_001125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.6, "hop_absolute_value": 2.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_clothes_episode_21-train-bf_000285-af_000255-minus-0000762", "task": "fold clothes", "image": [ "images/agilex_newdragon/fold_clothes/episode_11/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_clothes/episode_11/cam_high/frame_000844.jpg", "images/agilex_newdragon/fold_clothes/episode_21/cam_high/frame_000285.jpg", "images/agilex_newdragon/fold_clothes/episode_21/cam_left_wrist/frame_000285.jpg", "images/agilex_newdragon/fold_clothes/episode_21/cam_right_wrist/frame_000285.jpg", "images/agilex_newdragon/fold_clothes/episode_21/cam_high/frame_000255.jpg", "images/agilex_newdragon/fold_clothes/episode_21/cam_left_wrist/frame_000255.jpg", "images/agilex_newdragon/fold_clothes/episode_21/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.5, "hop_absolute_value": 10.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_15-train-bf_001185-af_000810-minus-0000263", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_15/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_15/cam_high/frame_001310.jpg", "images/agilex_newdragon/pour_coffee/episode_15/cam_high/frame_001185.jpg", "images/agilex_newdragon/pour_coffee/episode_15/cam_left_wrist/frame_001185.jpg", "images/agilex_newdragon/pour_coffee/episode_15/cam_right_wrist/frame_001185.jpg", "images/agilex_newdragon/pour_coffee/episode_15/cam_high/frame_000810.jpg", "images/agilex_newdragon/pour_coffee/episode_15/cam_left_wrist/frame_000810.jpg", "images/agilex_newdragon/pour_coffee/episode_15/cam_right_wrist/frame_000810.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -31.6, "hop_absolute_value": 31.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_21-train-bf_000885-af_000855-minus-0000633", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_24/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_24/cam_high/frame_001192.jpg", "images/agilex_newdragon/pour_coffee/episode_21/cam_high/frame_000885.jpg", "images/agilex_newdragon/pour_coffee/episode_21/cam_left_wrist/frame_000885.jpg", "images/agilex_newdragon/pour_coffee/episode_21/cam_right_wrist/frame_000885.jpg", "images/agilex_newdragon/pour_coffee/episode_21/cam_high/frame_000855.jpg", "images/agilex_newdragon/pour_coffee/episode_21/cam_left_wrist/frame_000855.jpg", "images/agilex_newdragon/pour_coffee/episode_21/cam_right_wrist/frame_000855.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.4, "hop_absolute_value": 3.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_9-train-bf_000390-af_000105-minus-0000121", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_9/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_9/cam_high/frame_000573.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_9/cam_high/frame_000390.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_9/cam_left_wrist/frame_000390.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_9/cam_right_wrist/frame_000390.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_9/cam_high/frame_000105.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_9/cam_left_wrist/frame_000105.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_9/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.1, "hop_absolute_value": 73.1, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_3-train-bf_000990-af_000195-minus-0000098", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_0/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_0/cam_high/frame_001397.jpg", "images/agilex_newdragon/pour_water/episode_3/cam_high/frame_000990.jpg", "images/agilex_newdragon/pour_water/episode_3/cam_left_wrist/frame_000990.jpg", "images/agilex_newdragon/pour_water/episode_3/cam_right_wrist/frame_000990.jpg", "images/agilex_newdragon/pour_water/episode_3/cam_high/frame_000195.jpg", "images/agilex_newdragon/pour_water/episode_3/cam_left_wrist/frame_000195.jpg", "images/agilex_newdragon/pour_water/episode_3/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -80.3, "hop_absolute_value": 80.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_12-train-bf_001125-af_001080-minus-0000755", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_5/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_5/cam_high/frame_001521.jpg", "images/agilex_newdragon/pour_coffee/episode_12/cam_high/frame_001125.jpg", "images/agilex_newdragon/pour_coffee/episode_12/cam_left_wrist/frame_001125.jpg", "images/agilex_newdragon/pour_coffee/episode_12/cam_right_wrist/frame_001125.jpg", "images/agilex_newdragon/pour_coffee/episode_12/cam_high/frame_001080.jpg", "images/agilex_newdragon/pour_coffee/episode_12/cam_left_wrist/frame_001080.jpg", "images/agilex_newdragon/pour_coffee/episode_12/cam_right_wrist/frame_001080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.0, "hop_absolute_value": 4.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_8-train-bf_000690-af_000585-minus-0000872", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_0/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_0/cam_high/frame_000827.jpg", "images/agilex_newdragon/glasses_storage/episode_8/cam_high/frame_000690.jpg", "images/agilex_newdragon/glasses_storage/episode_8/cam_left_wrist/frame_000690.jpg", "images/agilex_newdragon/glasses_storage/episode_8/cam_right_wrist/frame_000690.jpg", "images/agilex_newdragon/glasses_storage/episode_8/cam_high/frame_000585.jpg", "images/agilex_newdragon/glasses_storage/episode_8/cam_left_wrist/frame_000585.jpg", "images/agilex_newdragon/glasses_storage/episode_8/cam_right_wrist/frame_000585.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -15.2, "hop_absolute_value": 15.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_21-train-bf_000225-af_000000-minus-0000017", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_high/frame_000558.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_high/frame_000225.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_left_wrist/frame_000225.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_right_wrist/frame_000225.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_left_wrist/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_21-train-bf_000330-af_000285-minus-0000265", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_14/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_14/cam_high/frame_001261.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_high/frame_000330.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_left_wrist/frame_000330.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_right_wrist/frame_000330.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_high/frame_000285.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_left_wrist/frame_000285.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -13.6, "hop_absolute_value": 13.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_23-train-bf_001005-af_000435-minus-0000191", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_14/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_14/cam_high/frame_001771.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_high/frame_001005.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_left_wrist/frame_001005.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_right_wrist/frame_001005.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_high/frame_000435.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_left_wrist/frame_000435.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_right_wrist/frame_000435.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -56.7, "hop_absolute_value": 56.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_21-train-bf_000375-af_000360-minus-0000589", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_3/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_3/cam_high/frame_001245.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_high/frame_000375.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_left_wrist/frame_000375.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_right_wrist/frame_000375.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_high/frame_000360.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_left_wrist/frame_000360.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.0, "hop_absolute_value": 4.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_11-train-bf_001080-af_000330-minus-0000137", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_14/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_14/cam_high/frame_001771.jpg", "images/agilex_newdragon/clean_the_desk/episode_11/cam_high/frame_001080.jpg", "images/agilex_newdragon/clean_the_desk/episode_11/cam_left_wrist/frame_001080.jpg", "images/agilex_newdragon/clean_the_desk/episode_11/cam_right_wrist/frame_001080.jpg", "images/agilex_newdragon/clean_the_desk/episode_11/cam_high/frame_000330.jpg", "images/agilex_newdragon/clean_the_desk/episode_11/cam_left_wrist/frame_000330.jpg", "images/agilex_newdragon/clean_the_desk/episode_11/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -69.4, "hop_absolute_value": 69.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_17-train-bf_000630-af_000150-minus-0000089", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_0/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_0/cam_high/frame_001399.jpg", "images/agilex_newdragon/pour_coffee/episode_17/cam_high/frame_000630.jpg", "images/agilex_newdragon/pour_coffee/episode_17/cam_left_wrist/frame_000630.jpg", "images/agilex_newdragon/pour_coffee/episode_17/cam_right_wrist/frame_000630.jpg", "images/agilex_newdragon/pour_coffee/episode_17/cam_high/frame_000150.jpg", "images/agilex_newdragon/pour_coffee/episode_17/cam_left_wrist/frame_000150.jpg", "images/agilex_newdragon/pour_coffee/episode_17/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -76.2, "hop_absolute_value": 76.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_clothes_episode_16-train-bf_000645-af_000630-minus-0000675", "task": "fold clothes", "image": [ "images/agilex_newdragon/fold_clothes/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_clothes/episode_22/cam_high/frame_000896.jpg", "images/agilex_newdragon/fold_clothes/episode_16/cam_high/frame_000645.jpg", "images/agilex_newdragon/fold_clothes/episode_16/cam_left_wrist/frame_000645.jpg", "images/agilex_newdragon/fold_clothes/episode_16/cam_right_wrist/frame_000645.jpg", "images/agilex_newdragon/fold_clothes/episode_16/cam_high/frame_000630.jpg", "images/agilex_newdragon/fold_clothes/episode_16/cam_left_wrist/frame_000630.jpg", "images/agilex_newdragon/fold_clothes/episode_16/cam_right_wrist/frame_000630.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.3, "hop_absolute_value": 2.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_9-train-bf_001950-af_001890-minus-0000792", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_9/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_9/cam_high/frame_002381.jpg", "images/agilex_newdragon/fold_the_clothes/episode_9/cam_high/frame_001950.jpg", "images/agilex_newdragon/fold_the_clothes/episode_9/cam_left_wrist/frame_001950.jpg", "images/agilex_newdragon/fold_the_clothes/episode_9/cam_right_wrist/frame_001950.jpg", "images/agilex_newdragon/fold_the_clothes/episode_9/cam_high/frame_001890.jpg", "images/agilex_newdragon/fold_the_clothes/episode_9/cam_left_wrist/frame_001890.jpg", "images/agilex_newdragon/fold_the_clothes/episode_9/cam_right_wrist/frame_001890.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.1, "hop_absolute_value": 3.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_5-train-bf_000795-af_000720-minus-0000778", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_1/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_1/cam_high/frame_000984.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_high/frame_000795.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_left_wrist/frame_000795.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_right_wrist/frame_000795.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_high/frame_000720.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_left_wrist/frame_000720.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_right_wrist/frame_000720.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -9.4, "hop_absolute_value": 9.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_2-train-bf_000405-af_000390-minus-0000275", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_14/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_14/cam_high/frame_000572.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_2/cam_high/frame_000405.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_2/cam_left_wrist/frame_000405.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_2/cam_right_wrist/frame_000405.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_2/cam_high/frame_000390.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_2/cam_left_wrist/frame_000390.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_2/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.7, "hop_absolute_value": 3.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_22-train-bf_000450-af_000285-minus-0000932", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_21/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_21/cam_high/frame_000558.jpg", "images/agilex_newdragon/pour_water/episode_22/cam_high/frame_000450.jpg", "images/agilex_newdragon/pour_water/episode_22/cam_left_wrist/frame_000450.jpg", "images/agilex_newdragon/pour_water/episode_22/cam_right_wrist/frame_000450.jpg", "images/agilex_newdragon/pour_water/episode_22/cam_high/frame_000285.jpg", "images/agilex_newdragon/pour_water/episode_22/cam_left_wrist/frame_000285.jpg", "images/agilex_newdragon/pour_water/episode_22/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -36.7, "hop_absolute_value": 36.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_plug_insertion_episode_19-train-bf_000270-af_000135-minus-0000189", "task": "plug insertion", "image": [ "images/agilex_newdragon/plug_insertion/episode_14/cam_high/frame_000000.jpg", "images/agilex_newdragon/plug_insertion/episode_14/cam_high/frame_000508.jpg", "images/agilex_newdragon/plug_insertion/episode_19/cam_high/frame_000270.jpg", "images/agilex_newdragon/plug_insertion/episode_19/cam_left_wrist/frame_000270.jpg", "images/agilex_newdragon/plug_insertion/episode_19/cam_right_wrist/frame_000270.jpg", "images/agilex_newdragon/plug_insertion/episode_19/cam_high/frame_000135.jpg", "images/agilex_newdragon/plug_insertion/episode_19/cam_left_wrist/frame_000135.jpg", "images/agilex_newdragon/plug_insertion/episode_19/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`plug insertion`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `plug insertion`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_14-train-bf_001380-af_000060-minus-0000018", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_19/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_19/cam_high/frame_001671.jpg", "images/agilex_newdragon/clean_the_desk/episode_14/cam_high/frame_001380.jpg", "images/agilex_newdragon/clean_the_desk/episode_14/cam_left_wrist/frame_001380.jpg", "images/agilex_newdragon/clean_the_desk/episode_14/cam_right_wrist/frame_001380.jpg", "images/agilex_newdragon/clean_the_desk/episode_14/cam_high/frame_000060.jpg", "images/agilex_newdragon/clean_the_desk/episode_14/cam_left_wrist/frame_000060.jpg", "images/agilex_newdragon/clean_the_desk/episode_14/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -95.7, "hop_absolute_value": 95.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_15-train-bf_001050-af_000945-minus-0000903", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_23/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_23/cam_high/frame_001140.jpg", "images/agilex_newdragon/glasses_storage/episode_15/cam_high/frame_001050.jpg", "images/agilex_newdragon/glasses_storage/episode_15/cam_left_wrist/frame_001050.jpg", "images/agilex_newdragon/glasses_storage/episode_15/cam_right_wrist/frame_001050.jpg", "images/agilex_newdragon/glasses_storage/episode_15/cam_high/frame_000945.jpg", "images/agilex_newdragon/glasses_storage/episode_15/cam_left_wrist/frame_000945.jpg", "images/agilex_newdragon/glasses_storage/episode_15/cam_right_wrist/frame_000945.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_13-train-bf_000615-af_000015-minus-0000017", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_7/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_7/cam_high/frame_001270.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_high/frame_000615.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_left_wrist/frame_000615.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_right_wrist/frame_000615.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_high/frame_000015.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_left_wrist/frame_000015.jpg", "images/agilex_newdragon/clean_the_desk/episode_13/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -97.6, "hop_absolute_value": 97.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_6-train-bf_000915-af_000450-minus-0000221", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_3/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_3/cam_high/frame_001517.jpg", "images/agilex_newdragon/clean_the_desk/episode_6/cam_high/frame_000915.jpg", "images/agilex_newdragon/clean_the_desk/episode_6/cam_left_wrist/frame_000915.jpg", "images/agilex_newdragon/clean_the_desk/episode_6/cam_right_wrist/frame_000915.jpg", "images/agilex_newdragon/clean_the_desk/episode_6/cam_high/frame_000450.jpg", "images/agilex_newdragon/clean_the_desk/episode_6/cam_left_wrist/frame_000450.jpg", "images/agilex_newdragon/clean_the_desk/episode_6/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.8, "hop_absolute_value": 50.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_5-train-bf_000390-af_000240-minus-0000890", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_4/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_4/cam_high/frame_000487.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_5/cam_high/frame_000390.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_5/cam_left_wrist/frame_000390.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_5/cam_right_wrist/frame_000390.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_5/cam_high/frame_000240.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_5/cam_left_wrist/frame_000240.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_5/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -38.5, "hop_absolute_value": 38.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_20-train-bf_000630-af_000570-minus-0000906", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_23/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_23/cam_high/frame_000540.jpg", "images/agilex_newdragon/unplug/episode_20/cam_high/frame_000630.jpg", "images/agilex_newdragon/unplug/episode_20/cam_left_wrist/frame_000630.jpg", "images/agilex_newdragon/unplug/episode_20/cam_right_wrist/frame_000630.jpg", "images/agilex_newdragon/unplug/episode_20/cam_high/frame_000570.jpg", "images/agilex_newdragon/unplug/episode_20/cam_left_wrist/frame_000570.jpg", "images/agilex_newdragon/unplug/episode_20/cam_right_wrist/frame_000570.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -9.5, "hop_absolute_value": 9.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_storage_episode_3-train-bf_000120-af_000015-minus-0000043", "task": "glasses storage", "image": [ "images/agilex_newdragon/glasses_storage/episode_3/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_high/frame_000643.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_high/frame_000120.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_left_wrist/frame_000120.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_right_wrist/frame_000120.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_high/frame_000015.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_left_wrist/frame_000015.jpg", "images/agilex_newdragon/glasses_storage/episode_3/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_glasses_case_episode_15-train-bf_000045-af_000030-minus-0000624", "task": "glasses case", "image": [ "images/agilex_newdragon/glasses_case/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/glasses_case/episode_13/cam_high/frame_000622.jpg", "images/agilex_newdragon/glasses_case/episode_15/cam_high/frame_000045.jpg", "images/agilex_newdragon/glasses_case/episode_15/cam_left_wrist/frame_000045.jpg", "images/agilex_newdragon/glasses_case/episode_15/cam_right_wrist/frame_000045.jpg", "images/agilex_newdragon/glasses_case/episode_15/cam_high/frame_000030.jpg", "images/agilex_newdragon/glasses_case/episode_15/cam_left_wrist/frame_000030.jpg", "images/agilex_newdragon/glasses_case/episode_15/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`glasses case`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `glasses case`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_0-train-bf_001875-af_000405-minus-0000100", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_20/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_20/cam_high/frame_002048.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_high/frame_001875.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_left_wrist/frame_001875.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_right_wrist/frame_001875.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_high/frame_000405.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_left_wrist/frame_000405.jpg", "images/agilex_newdragon/fold_the_clothes/episode_0/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -78.4, "hop_absolute_value": 78.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_into_the_cup_episode_10-train-bf_000555-af_000495-minus-0000656", "task": "pour water into the cup", "image": [ "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_high/frame_000667.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_high/frame_000555.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_left_wrist/frame_000555.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_right_wrist/frame_000555.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_high/frame_000495.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_left_wrist/frame_000495.jpg", "images/agilex_newdragon/pour_water_into_the_cup/episode_10/cam_right_wrist/frame_000495.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water into the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water into the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.8, "hop_absolute_value": 10.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_water_episode_13-train-bf_000780-af_000750-minus-0000634", "task": "pour water", "image": [ "images/agilex_newdragon/pour_water/episode_18/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_water/episode_18/cam_high/frame_001415.jpg", "images/agilex_newdragon/pour_water/episode_13/cam_high/frame_000780.jpg", "images/agilex_newdragon/pour_water/episode_13/cam_left_wrist/frame_000780.jpg", "images/agilex_newdragon/pour_water/episode_13/cam_right_wrist/frame_000780.jpg", "images/agilex_newdragon/pour_water/episode_13/cam_high/frame_000750.jpg", "images/agilex_newdragon/pour_water/episode_13/cam_left_wrist/frame_000750.jpg", "images/agilex_newdragon/pour_water/episode_13/cam_right_wrist/frame_000750.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.8, "hop_absolute_value": 3.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_20-train-bf_000615-af_000045-minus-0000011", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_5/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_5/cam_high/frame_001845.jpg", "images/agilex_newdragon/fold_the_clothes/episode_20/cam_high/frame_000615.jpg", "images/agilex_newdragon/fold_the_clothes/episode_20/cam_left_wrist/frame_000615.jpg", "images/agilex_newdragon/fold_the_clothes/episode_20/cam_right_wrist/frame_000615.jpg", "images/agilex_newdragon/fold_the_clothes/episode_20/cam_high/frame_000045.jpg", "images/agilex_newdragon/fold_the_clothes/episode_20/cam_left_wrist/frame_000045.jpg", "images/agilex_newdragon/fold_the_clothes/episode_20/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -92.7, "hop_absolute_value": 92.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_unplug_episode_13-train-bf_000345-af_000165-minus-0000842", "task": "unplug", "image": [ "images/agilex_newdragon/unplug/episode_0/cam_high/frame_000000.jpg", "images/agilex_newdragon/unplug/episode_0/cam_high/frame_000584.jpg", "images/agilex_newdragon/unplug/episode_13/cam_high/frame_000345.jpg", "images/agilex_newdragon/unplug/episode_13/cam_left_wrist/frame_000345.jpg", "images/agilex_newdragon/unplug/episode_13/cam_right_wrist/frame_000345.jpg", "images/agilex_newdragon/unplug/episode_13/cam_high/frame_000165.jpg", "images/agilex_newdragon/unplug/episode_13/cam_left_wrist/frame_000165.jpg", "images/agilex_newdragon/unplug/episode_13/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`unplug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `unplug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -52.2, "hop_absolute_value": 52.2, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_23-train-bf_000300-af_000000-minus-0000008", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_23/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_high/frame_001845.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_high/frame_000300.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_left_wrist/frame_000300.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_right_wrist/frame_000300.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_left_wrist/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_23/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_18-train-bf_000780-af_000765-minus-0000818", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_9/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_9/cam_high/frame_001398.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_high/frame_000780.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_left_wrist/frame_000780.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_right_wrist/frame_000780.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_high/frame_000765.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_left_wrist/frame_000765.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_right_wrist/frame_000765.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.9, "hop_absolute_value": 1.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_3-train-bf_001050-af_001035-minus-0000897", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_18/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_18/cam_high/frame_001768.jpg", "images/agilex_newdragon/clean_the_desk/episode_3/cam_high/frame_001050.jpg", "images/agilex_newdragon/clean_the_desk/episode_3/cam_left_wrist/frame_001050.jpg", "images/agilex_newdragon/clean_the_desk/episode_3/cam_right_wrist/frame_001050.jpg", "images/agilex_newdragon/clean_the_desk/episode_3/cam_high/frame_001035.jpg", "images/agilex_newdragon/clean_the_desk/episode_3/cam_left_wrist/frame_001035.jpg", "images/agilex_newdragon/clean_the_desk/episode_3/cam_right_wrist/frame_001035.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.4, "hop_absolute_value": 1.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_21-train-bf_001035-af_001020-minus-0000880", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_8/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_8/cam_high/frame_002314.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_high/frame_001035.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_left_wrist/frame_001035.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_right_wrist/frame_001035.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_high/frame_001020.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_left_wrist/frame_001020.jpg", "images/agilex_newdragon/fold_the_clothes/episode_21/cam_right_wrist/frame_001020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.4, "hop_absolute_value": 1.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_7-train-bf_000690-af_000675-minus-0000656", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_24/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_24/cam_high/frame_001192.jpg", "images/agilex_newdragon/pour_coffee/episode_7/cam_high/frame_000690.jpg", "images/agilex_newdragon/pour_coffee/episode_7/cam_left_wrist/frame_000690.jpg", "images/agilex_newdragon/pour_coffee/episode_7/cam_right_wrist/frame_000690.jpg", "images/agilex_newdragon/pour_coffee/episode_7/cam_high/frame_000675.jpg", "images/agilex_newdragon/pour_coffee/episode_7/cam_left_wrist/frame_000675.jpg", "images/agilex_newdragon/pour_coffee/episode_7/cam_right_wrist/frame_000675.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.2, "hop_absolute_value": 2.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_clean_the_desk_episode_11-train-bf_001410-af_001395-minus-0000837", "task": "clean the desk", "image": [ "images/agilex_newdragon/clean_the_desk/episode_2/cam_high/frame_000000.jpg", "images/agilex_newdragon/clean_the_desk/episode_2/cam_high/frame_000954.jpg", "images/agilex_newdragon/clean_the_desk/episode_11/cam_high/frame_001410.jpg", "images/agilex_newdragon/clean_the_desk/episode_11/cam_left_wrist/frame_001410.jpg", "images/agilex_newdragon/clean_the_desk/episode_11/cam_right_wrist/frame_001410.jpg", "images/agilex_newdragon/clean_the_desk/episode_11/cam_high/frame_001395.jpg", "images/agilex_newdragon/clean_the_desk/episode_11/cam_left_wrist/frame_001395.jpg", "images/agilex_newdragon/clean_the_desk/episode_11/cam_right_wrist/frame_001395.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`clean the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `clean the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.1, "hop_absolute_value": 1.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_2-train-bf_000420-af_000405-minus-0000866", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_003220.jpg", "images/agilex_newdragon/fold_the_clothes/episode_2/cam_high/frame_000420.jpg", "images/agilex_newdragon/fold_the_clothes/episode_2/cam_left_wrist/frame_000420.jpg", "images/agilex_newdragon/fold_the_clothes/episode_2/cam_right_wrist/frame_000420.jpg", "images/agilex_newdragon/fold_the_clothes/episode_2/cam_high/frame_000405.jpg", "images/agilex_newdragon/fold_the_clothes/episode_2/cam_left_wrist/frame_000405.jpg", "images/agilex_newdragon/fold_the_clothes/episode_2/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.6, "hop_absolute_value": 3.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_13-train-bf_000240-af_000225-minus-0000641", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_19/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_19/cam_high/frame_000589.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_high/frame_000240.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_left_wrist/frame_000240.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_right_wrist/frame_000240.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_high/frame_000225.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_left_wrist/frame_000225.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.2, "hop_absolute_value": 6.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_5-train-bf_000750-af_000735-minus-0000639", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_19/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_19/cam_high/frame_001456.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_high/frame_000750.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_left_wrist/frame_000750.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_right_wrist/frame_000750.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_high/frame_000735.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_left_wrist/frame_000735.jpg", "images/agilex_newdragon/food_packaging/episode_5/cam_right_wrist/frame_000735.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.0, "hop_absolute_value": 2.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_6-train-bf_001080-af_001050-minus-0000886", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_003220.jpg", "images/agilex_newdragon/fold_the_clothes/episode_6/cam_high/frame_001080.jpg", "images/agilex_newdragon/fold_the_clothes/episode_6/cam_left_wrist/frame_001080.jpg", "images/agilex_newdragon/fold_the_clothes/episode_6/cam_right_wrist/frame_001080.jpg", "images/agilex_newdragon/fold_the_clothes/episode_6/cam_high/frame_001050.jpg", "images/agilex_newdragon/fold_the_clothes/episode_6/cam_left_wrist/frame_001050.jpg", "images/agilex_newdragon/fold_the_clothes/episode_6/cam_right_wrist/frame_001050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.8, "hop_absolute_value": 2.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_pour_coffee_episode_3-train-bf_000960-af_000930-minus-0000681", "task": "pour coffee", "image": [ "images/agilex_newdragon/pour_coffee/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/pour_coffee/episode_22/cam_high/frame_001136.jpg", "images/agilex_newdragon/pour_coffee/episode_3/cam_high/frame_000960.jpg", "images/agilex_newdragon/pour_coffee/episode_3/cam_left_wrist/frame_000960.jpg", "images/agilex_newdragon/pour_coffee/episode_3/cam_right_wrist/frame_000960.jpg", "images/agilex_newdragon/pour_coffee/episode_3/cam_high/frame_000930.jpg", "images/agilex_newdragon/pour_coffee/episode_3/cam_left_wrist/frame_000930.jpg", "images/agilex_newdragon/pour_coffee/episode_3/cam_right_wrist/frame_000930.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.1, "hop_absolute_value": 3.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_23-train-bf_000750-af_000720-minus-0000633", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_003220.jpg", "images/agilex_newdragon/fold_the_clothes/episode_23/cam_high/frame_000750.jpg", "images/agilex_newdragon/fold_the_clothes/episode_23/cam_left_wrist/frame_000750.jpg", "images/agilex_newdragon/fold_the_clothes/episode_23/cam_right_wrist/frame_000750.jpg", "images/agilex_newdragon/fold_the_clothes/episode_23/cam_high/frame_000720.jpg", "images/agilex_newdragon/fold_the_clothes/episode_23/cam_left_wrist/frame_000720.jpg", "images/agilex_newdragon/fold_the_clothes/episode_23/cam_right_wrist/frame_000720.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.0, "hop_absolute_value": 4.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_fold_the_clothes_episode_11-train-bf_002370-af_002295-minus-0000834", "task": "fold the clothes", "image": [ "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_000000.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_003220.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_002370.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_left_wrist/frame_002370.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_right_wrist/frame_002370.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_high/frame_002295.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_left_wrist/frame_002295.jpg", "images/agilex_newdragon/fold_the_clothes/episode_11/cam_right_wrist/frame_002295.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold the clothes`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold the clothes`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.2, "hop_absolute_value": 3.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_20-train-bf_001155-af_001140-minus-0000774", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_20/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_high/frame_001461.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_high/frame_001155.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_left_wrist/frame_001155.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_right_wrist/frame_001155.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_high/frame_001140.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_left_wrist/frame_001140.jpg", "images/agilex_newdragon/food_packaging/episode_20/cam_right_wrist/frame_001140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.3, "hop_absolute_value": 1.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_22-train-bf_000360-af_000165-minus-0000176", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_high/frame_000544.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_high/frame_000360.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_left_wrist/frame_000360.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_right_wrist/frame_000360.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_high/frame_000165.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_left_wrist/frame_000165.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_22/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -54.2, "hop_absolute_value": 54.2, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_twist_the_bottle_cap_episode_19-train-bf_000375-af_000330-minus-0000614", "task": "twist the bottle cap", "image": [ "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_high/frame_000000.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_13/cam_high/frame_000618.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_19/cam_high/frame_000375.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_19/cam_left_wrist/frame_000375.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_19/cam_right_wrist/frame_000375.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_19/cam_high/frame_000330.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_19/cam_left_wrist/frame_000330.jpg", "images/agilex_newdragon/twist_the_bottle_cap/episode_19/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`twist the bottle cap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `twist the bottle cap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.0, "hop_absolute_value": 12.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "agilex_newdragon_food_packaging_episode_22-train-bf_000855-af_000825-minus-0000833", "task": "food packaging", "image": [ "images/agilex_newdragon/food_packaging/episode_0/cam_high/frame_000000.jpg", "images/agilex_newdragon/food_packaging/episode_0/cam_high/frame_000999.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_high/frame_000855.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_left_wrist/frame_000855.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_right_wrist/frame_000855.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_high/frame_000825.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_left_wrist/frame_000825.jpg", "images/agilex_newdragon/food_packaging/episode_22/cam_right_wrist/frame_000825.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`food packaging`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `food packaging`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.5, "hop_absolute_value": 3.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "agilex_newdragon" }, { "id": "droid_oxe_Put_one_orange_chip_in_the_blue_gameboard_episode_3-train-bf_000015-af_000270-plus-0000387", "task": "Put one orange chip in the blue gameboard", "image": [ "images/droid_oxe/Put_one_orange_chip_in_the_blue_gameboard/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_one_orange_chip_in_the_blue_gameboard/episode_0/cam_high/frame_000397.jpg", "images/droid_oxe/Put_one_orange_chip_in_the_blue_gameboard/episode_3/cam_high/frame_000015.jpg", "images/droid_oxe/Put_one_orange_chip_in_the_blue_gameboard/episode_3/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Put_one_orange_chip_in_the_blue_gameboard/episode_3/cam_right_wrist/frame_000015.jpg", "images/droid_oxe/Put_one_orange_chip_in_the_blue_gameboard/episode_3/cam_high/frame_000270.jpg", "images/droid_oxe/Put_one_orange_chip_in_the_blue_gameboard/episode_3/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Put_one_orange_chip_in_the_blue_gameboard/episode_3/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put one orange chip in the blue gameboard`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put one orange chip in the blue gameboard`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.8, "hop_absolute_value": 70.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_black_marker_from_the_white_bowl_episode_0-train-bf_000045-af_000090-plus-0000113", "task": "Remove the black marker from the white bowl", "image": [ "images/droid_oxe/Remove_the_black_marker_from_the_white_bowl/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_black_marker_from_the_white_bowl/episode_0/cam_high/frame_000196.jpg", "images/droid_oxe/Remove_the_black_marker_from_the_white_bowl/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Remove_the_black_marker_from_the_white_bowl/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Remove_the_black_marker_from_the_white_bowl/episode_0/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Remove_the_black_marker_from_the_white_bowl/episode_0/cam_high/frame_000090.jpg", "images/droid_oxe/Remove_the_black_marker_from_the_white_bowl/episode_0/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Remove_the_black_marker_from_the_white_bowl/episode_0/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the black marker from the white bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the black marker from the white bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 30.0, "hop_absolute_value": 30.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Turn_on_the_switch_on_the_adapter_and_then_close_the_clear_container_episode_0-train-bf_000075-af_000235-plus-0000228", "task": "Turn on the switch on the adapter and then close the clear container", "image": [ "images/droid_oxe/Turn_on_the_switch_on_the_adapter_and_then_close_the_clear_container/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Turn_on_the_switch_on_the_adapter_and_then_close_the_clear_container/episode_0/cam_high/frame_000235.jpg", "images/droid_oxe/Turn_on_the_switch_on_the_adapter_and_then_close_the_clear_container/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Turn_on_the_switch_on_the_adapter_and_then_close_the_clear_container/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Turn_on_the_switch_on_the_adapter_and_then_close_the_clear_container/episode_0/cam_right_wrist/frame_000075.jpg", "images/droid_oxe/Turn_on_the_switch_on_the_adapter_and_then_close_the_clear_container/episode_0/cam_high/frame_000235.jpg", "images/droid_oxe/Turn_on_the_switch_on_the_adapter_and_then_close_the_clear_container/episode_0/cam_left_wrist/frame_000235.jpg", "images/droid_oxe/Turn_on_the_switch_on_the_adapter_and_then_close_the_clear_container/episode_0/cam_right_wrist/frame_000235.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Turn on the switch on the adapter and then close the clear container`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Turn on the switch on the adapter and then close the clear container`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_marker_in_the_teal_bowl_episode_0-train-bf_000015-af_000165-plus-0000239", "task": "Put the marker in the teal bowl", "image": [ "images/droid_oxe/Put_the_marker_in_the_teal_bowl/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_marker_in_the_teal_bowl/episode_0/cam_high/frame_000261.jpg", "images/droid_oxe/Put_the_marker_in_the_teal_bowl/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Put_the_marker_in_the_teal_bowl/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Put_the_marker_in_the_teal_bowl/episode_0/cam_right_wrist/frame_000015.jpg", "images/droid_oxe/Put_the_marker_in_the_teal_bowl/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Put_the_marker_in_the_teal_bowl/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Put_the_marker_in_the_teal_bowl/episode_0/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the marker in the teal bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the marker in the teal bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Transfer_the_lid_from_the_black_pot_to_the_blue_pot_episode_2-train-bf_000105-af_000285-plus-0000709", "task": "Transfer the lid from the black pot to the blue pot", "image": [ "images/droid_oxe/Transfer_the_lid_from_the_black_pot_to_the_blue_pot/episode_2/cam_high/frame_000000.jpg", "images/droid_oxe/Transfer_the_lid_from_the_black_pot_to_the_blue_pot/episode_2/cam_high/frame_000406.jpg", "images/droid_oxe/Transfer_the_lid_from_the_black_pot_to_the_blue_pot/episode_2/cam_high/frame_000105.jpg", "images/droid_oxe/Transfer_the_lid_from_the_black_pot_to_the_blue_pot/episode_2/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Transfer_the_lid_from_the_black_pot_to_the_blue_pot/episode_2/cam_right_wrist/frame_000105.jpg", "images/droid_oxe/Transfer_the_lid_from_the_black_pot_to_the_blue_pot/episode_2/cam_high/frame_000285.jpg", "images/droid_oxe/Transfer_the_lid_from_the_black_pot_to_the_blue_pot/episode_2/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/Transfer_the_lid_from_the_black_pot_to_the_blue_pot/episode_2/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Transfer the lid from the black pot to the blue pot`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Transfer the lid from the black pot to the blue pot`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_yellow_block_on_the_right_and_place_it_on_top_of_the_green_block_episode_0-train-bf_000075-af_000180-plus-0000196", "task": "Pick up the yellow block on the right and place it on top of the green block", "image": [ "images/droid_oxe/Pick_up_the_yellow_block_on_the_right_and_place_it_on_top_of_the_green_block/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_yellow_block_on_the_right_and_place_it_on_top_of_the_green_block/episode_0/cam_high/frame_000236.jpg", "images/droid_oxe/Pick_up_the_yellow_block_on_the_right_and_place_it_on_top_of_the_green_block/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Pick_up_the_yellow_block_on_the_right_and_place_it_on_top_of_the_green_block/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Pick_up_the_yellow_block_on_the_right_and_place_it_on_top_of_the_green_block/episode_0/cam_right_wrist/frame_000075.jpg", "images/droid_oxe/Pick_up_the_yellow_block_on_the_right_and_place_it_on_top_of_the_green_block/episode_0/cam_high/frame_000180.jpg", "images/droid_oxe/Pick_up_the_yellow_block_on_the_right_and_place_it_on_top_of_the_green_block/episode_0/cam_left_wrist/frame_000180.jpg", "images/droid_oxe/Pick_up_the_yellow_block_on_the_right_and_place_it_on_top_of_the_green_block/episode_0/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the yellow block on the right and place it on top of the green block`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the yellow block on the right and place it on top of the green block`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_chopstick_to_stir_in_the_transparent_bowl_episode_0-train-bf_000090-af_000180-plus-0000800", "task": "Use the chopstick to stir in the transparent bowl", "image": [ "images/droid_oxe/Use_the_chopstick_to_stir_in_the_transparent_bowl/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_chopstick_to_stir_in_the_transparent_bowl/episode_0/cam_high/frame_000537.jpg", "images/droid_oxe/Use_the_chopstick_to_stir_in_the_transparent_bowl/episode_0/cam_high/frame_000090.jpg", "images/droid_oxe/Use_the_chopstick_to_stir_in_the_transparent_bowl/episode_0/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Use_the_chopstick_to_stir_in_the_transparent_bowl/episode_0/cam_right_wrist/frame_000090.jpg", "images/droid_oxe/Use_the_chopstick_to_stir_in_the_transparent_bowl/episode_0/cam_high/frame_000180.jpg", "images/droid_oxe/Use_the_chopstick_to_stir_in_the_transparent_bowl/episode_0/cam_left_wrist/frame_000180.jpg", "images/droid_oxe/Use_the_chopstick_to_stir_in_the_transparent_bowl/episode_0/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the chopstick to stir in the transparent bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the chopstick to stir in the transparent bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.7, "hop_absolute_value": 20.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pour_the_things_in_the_green_cup_onto_the_plate_episode_2-train-bf_000150-af_000270-plus-0000341", "task": "Pour the things in the green cup onto the plate", "image": [ "images/droid_oxe/Pour_the_things_in_the_green_cup_onto_the_plate/episode_2/cam_high/frame_000000.jpg", "images/droid_oxe/Pour_the_things_in_the_green_cup_onto_the_plate/episode_2/cam_high/frame_000382.jpg", "images/droid_oxe/Pour_the_things_in_the_green_cup_onto_the_plate/episode_2/cam_high/frame_000150.jpg", "images/droid_oxe/Pour_the_things_in_the_green_cup_onto_the_plate/episode_2/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Pour_the_things_in_the_green_cup_onto_the_plate/episode_2/cam_right_wrist/frame_000150.jpg", "images/droid_oxe/Pour_the_things_in_the_green_cup_onto_the_plate/episode_2/cam_high/frame_000270.jpg", "images/droid_oxe/Pour_the_things_in_the_green_cup_onto_the_plate/episode_2/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Pour_the_things_in_the_green_cup_onto_the_plate/episode_2/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pour the things in the green cup onto the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pour the things in the green cup onto the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 53.3, "hop_absolute_value": 53.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_pick_up_the_lid_and_place_it_on_the_pot_episode_0-train-bf_000000-af_000150-plus-0000137", "task": "pick up the lid and place it on the pot", "image": [ "images/droid_oxe/pick_up_the_lid_and_place_it_on_the_pot/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/pick_up_the_lid_and_place_it_on_the_pot/episode_0/cam_high/frame_000180.jpg", "images/droid_oxe/pick_up_the_lid_and_place_it_on_the_pot/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/pick_up_the_lid_and_place_it_on_the_pot/episode_0/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/pick_up_the_lid_and_place_it_on_the_pot/episode_0/cam_right_wrist/frame_000000.jpg", "images/droid_oxe/pick_up_the_lid_and_place_it_on_the_pot/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/pick_up_the_lid_and_place_it_on_the_pot/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/pick_up_the_lid_and_place_it_on_the_pot/episode_0/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the lid and place it on the pot`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the lid and place it on the pot`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_lid_from_the_big_pot_episode_0-train-bf_000165-af_000225-plus-0000261", "task": "Remove the lid from the big pot", "image": [ "images/droid_oxe/Remove_the_lid_from_the_big_pot/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_lid_from_the_big_pot/episode_0/cam_high/frame_000335.jpg", "images/droid_oxe/Remove_the_lid_from_the_big_pot/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Remove_the_lid_from_the_big_pot/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Remove_the_lid_from_the_big_pot/episode_0/cam_right_wrist/frame_000165.jpg", "images/droid_oxe/Remove_the_lid_from_the_big_pot/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Remove_the_lid_from_the_big_pot/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Remove_the_lid_from_the_big_pot/episode_0/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the lid from the big pot`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the lid from the big pot`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 36.4, "hop_absolute_value": 36.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_green_bowl_to_the_left_side_of_the_table_remove_the_silver_lid_from_the_top_of_the_silver_tin_episode_0-train-bf_000165-af_000270-plus-0000657", "task": "Move the green bowl to the left side of the table remove the silver lid from the top of the silver tin", "image": [ "images/droid_oxe/Move_the_green_bowl_to_the_left_side_of_the_table_remove_the_silver_lid_from_the_top_of_the_silver_tin/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_green_bowl_to_the_left_side_of_the_table_remove_the_silver_lid_from_the_top_of_the_silver_tin/episode_0/cam_high/frame_000408.jpg", "images/droid_oxe/Move_the_green_bowl_to_the_left_side_of_the_table_remove_the_silver_lid_from_the_top_of_the_silver_tin/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Move_the_green_bowl_to_the_left_side_of_the_table_remove_the_silver_lid_from_the_top_of_the_silver_tin/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Move_the_green_bowl_to_the_left_side_of_the_table_remove_the_silver_lid_from_the_top_of_the_silver_tin/episode_0/cam_right_wrist/frame_000165.jpg", "images/droid_oxe/Move_the_green_bowl_to_the_left_side_of_the_table_remove_the_silver_lid_from_the_top_of_the_silver_tin/episode_0/cam_high/frame_000270.jpg", "images/droid_oxe/Move_the_green_bowl_to_the_left_side_of_the_table_remove_the_silver_lid_from_the_top_of_the_silver_tin/episode_0/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Move_the_green_bowl_to_the_left_side_of_the_table_remove_the_silver_lid_from_the_top_of_the_silver_tin/episode_0/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the green bowl to the left side of the table remove the silver lid from the top of the silver tin`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the green bowl to the left side of the table remove the silver lid from the top of the silver tin`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 43.8, "hop_absolute_value": 43.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_case_on_the_cloth_directly_on_the_table_episode_0-train-bf_000405-af_000555-plus-0000428", "task": "Put the case on the cloth directly on the table", "image": [ "images/droid_oxe/Put_the_case_on_the_cloth_directly_on_the_table/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_case_on_the_cloth_directly_on_the_table/episode_0/cam_high/frame_000651.jpg", "images/droid_oxe/Put_the_case_on_the_cloth_directly_on_the_table/episode_0/cam_high/frame_000405.jpg", "images/droid_oxe/Put_the_case_on_the_cloth_directly_on_the_table/episode_0/cam_left_wrist/frame_000405.jpg", "images/droid_oxe/Put_the_case_on_the_cloth_directly_on_the_table/episode_0/cam_right_wrist/frame_000405.jpg", "images/droid_oxe/Put_the_case_on_the_cloth_directly_on_the_table/episode_0/cam_high/frame_000555.jpg", "images/droid_oxe/Put_the_case_on_the_cloth_directly_on_the_table/episode_0/cam_left_wrist/frame_000555.jpg", "images/droid_oxe/Put_the_case_on_the_cloth_directly_on_the_table/episode_0/cam_right_wrist/frame_000555.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the case on the cloth directly on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the case on the cloth directly on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Cover_the_the_silver_pot_with_the_silver_lid_episode_0-train-bf_000015-af_000285-plus-0000370", "task": "Cover the the silver pot with the silver lid", "image": [ "images/droid_oxe/Cover_the_the_silver_pot_with_the_silver_lid/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Cover_the_the_silver_pot_with_the_silver_lid/episode_0/cam_high/frame_000318.jpg", "images/droid_oxe/Cover_the_the_silver_pot_with_the_silver_lid/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Cover_the_the_silver_pot_with_the_silver_lid/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Cover_the_the_silver_pot_with_the_silver_lid/episode_0/cam_right_wrist/frame_000015.jpg", "images/droid_oxe/Cover_the_the_silver_pot_with_the_silver_lid/episode_0/cam_high/frame_000285.jpg", "images/droid_oxe/Cover_the_the_silver_pot_with_the_silver_lid/episode_0/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/Cover_the_the_silver_pot_with_the_silver_lid/episode_0/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Cover the the silver pot with the silver lid`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Cover the the silver pot with the silver lid`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 90.0, "hop_absolute_value": 90.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_plate_in_the_dishwasher_on_the_plate_on_the_counter_episode_0-train-bf_000150-af_000270-plus-0000795", "task": "Put the plate in the dishwasher on the plate on the counter", "image": [ "images/droid_oxe/Put_the_plate_in_the_dishwasher_on_the_plate_on_the_counter/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_plate_in_the_dishwasher_on_the_plate_on_the_counter/episode_0/cam_high/frame_000476.jpg", "images/droid_oxe/Put_the_plate_in_the_dishwasher_on_the_plate_on_the_counter/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Put_the_plate_in_the_dishwasher_on_the_plate_on_the_counter/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Put_the_plate_in_the_dishwasher_on_the_plate_on_the_counter/episode_0/cam_right_wrist/frame_000150.jpg", "images/droid_oxe/Put_the_plate_in_the_dishwasher_on_the_plate_on_the_counter/episode_0/cam_high/frame_000270.jpg", "images/droid_oxe/Put_the_plate_in_the_dishwasher_on_the_plate_on_the_counter/episode_0/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Put_the_plate_in_the_dishwasher_on_the_plate_on_the_counter/episode_0/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the plate in the dishwasher on the plate on the counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the plate in the dishwasher on the plate on the counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 38.1, "hop_absolute_value": 38.1, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_orange_straw_in_the_can_episode_0-train-bf_000000-af_000105-plus-0000795", "task": "Put the orange straw in the can", "image": [ "images/droid_oxe/Put_the_orange_straw_in_the_can/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_orange_straw_in_the_can/episode_0/cam_high/frame_000532.jpg", "images/droid_oxe/Put_the_orange_straw_in_the_can/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_orange_straw_in_the_can/episode_0/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Put_the_orange_straw_in_the_can/episode_0/cam_right_wrist/frame_000000.jpg", "images/droid_oxe/Put_the_orange_straw_in_the_can/episode_0/cam_high/frame_000105.jpg", "images/droid_oxe/Put_the_orange_straw_in_the_can/episode_0/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Put_the_orange_straw_in_the_can/episode_0/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the orange straw in the can`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the orange straw in the can`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Press_the_top_of_the_liquid_soap_bottle_episode_0-train-bf_000075-af_000150-plus-0000127", "task": "Press the top of the liquid soap bottle", "image": [ "images/droid_oxe/Press_the_top_of_the_liquid_soap_bottle/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Press_the_top_of_the_liquid_soap_bottle/episode_0/cam_high/frame_000191.jpg", "images/droid_oxe/Press_the_top_of_the_liquid_soap_bottle/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Press_the_top_of_the_liquid_soap_bottle/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Press_the_top_of_the_liquid_soap_bottle/episode_0/cam_right_wrist/frame_000075.jpg", "images/droid_oxe/Press_the_top_of_the_liquid_soap_bottle/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Press_the_top_of_the_liquid_soap_bottle/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Press_the_top_of_the_liquid_soap_bottle/episode_0/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Press the top of the liquid soap bottle`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Press the top of the liquid soap bottle`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_red-white_book_and_the_purple_object_in_the_black_box_episode_0-train-bf_000150-af_000225-plus-0000648", "task": "Put the red-white book and the purple object in the black box", "image": [ "images/droid_oxe/Put_the_red-white_book_and_the_purple_object_in_the_black_box/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_red-white_book_and_the_purple_object_in_the_black_box/episode_0/cam_high/frame_000441.jpg", "images/droid_oxe/Put_the_red-white_book_and_the_purple_object_in_the_black_box/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Put_the_red-white_book_and_the_purple_object_in_the_black_box/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Put_the_red-white_book_and_the_purple_object_in_the_black_box/episode_0/cam_right_wrist/frame_000150.jpg", "images/droid_oxe/Put_the_red-white_book_and_the_purple_object_in_the_black_box/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Put_the_red-white_book_and_the_purple_object_in_the_black_box/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_red-white_book_and_the_purple_object_in_the_black_box/episode_0/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the red-white book and the purple object in the black box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the red-white book and the purple object in the black box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 26.3, "hop_absolute_value": 26.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_desk_lamp_to_the_right_then_take_the_rubber_duck_and_place_it_on_the_middle_shelf_episode_0-train-bf_000660-af_000893-plus-0000568", "task": "Move the desk lamp to the right then take the rubber duck and place it on the middle shelf", "image": [ "images/droid_oxe/Move_the_desk_lamp_to_the_right_then_take_the_rubber_duck_and_place_it_on_the_middle_shelf/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_desk_lamp_to_the_right_then_take_the_rubber_duck_and_place_it_on_the_middle_shelf/episode_0/cam_high/frame_000893.jpg", "images/droid_oxe/Move_the_desk_lamp_to_the_right_then_take_the_rubber_duck_and_place_it_on_the_middle_shelf/episode_0/cam_high/frame_000660.jpg", "images/droid_oxe/Move_the_desk_lamp_to_the_right_then_take_the_rubber_duck_and_place_it_on_the_middle_shelf/episode_0/cam_left_wrist/frame_000660.jpg", "images/droid_oxe/Move_the_desk_lamp_to_the_right_then_take_the_rubber_duck_and_place_it_on_the_middle_shelf/episode_0/cam_right_wrist/frame_000660.jpg", "images/droid_oxe/Move_the_desk_lamp_to_the_right_then_take_the_rubber_duck_and_place_it_on_the_middle_shelf/episode_0/cam_high/frame_000893.jpg", "images/droid_oxe/Move_the_desk_lamp_to_the_right_then_take_the_rubber_duck_and_place_it_on_the_middle_shelf/episode_0/cam_left_wrist/frame_000893.jpg", "images/droid_oxe/Move_the_desk_lamp_to_the_right_then_take_the_rubber_duck_and_place_it_on_the_middle_shelf/episode_0/cam_right_wrist/frame_000893.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the desk lamp to the right then take the rubber duck and place it on the middle shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the desk lamp to the right then take the rubber duck and place it on the middle shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_black_key_and_place_it_on_the_tray_on_the_left_episode_0-train-bf_000030-af_000075-plus-0000239", "task": "Pick up the black key and place it on the tray on the left", "image": [ "images/droid_oxe/Pick_up_the_black_key_and_place_it_on_the_tray_on_the_left/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_black_key_and_place_it_on_the_tray_on_the_left/episode_0/cam_high/frame_000229.jpg", "images/droid_oxe/Pick_up_the_black_key_and_place_it_on_the_tray_on_the_left/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Pick_up_the_black_key_and_place_it_on_the_tray_on_the_left/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Pick_up_the_black_key_and_place_it_on_the_tray_on_the_left/episode_0/cam_right_wrist/frame_000030.jpg", "images/droid_oxe/Pick_up_the_black_key_and_place_it_on_the_tray_on_the_left/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Pick_up_the_black_key_and_place_it_on_the_tray_on_the_left/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Pick_up_the_black_key_and_place_it_on_the_tray_on_the_left/episode_0/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the black key and place it on the tray on the left`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the black key and place it on the tray on the left`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 23.1, "hop_absolute_value": 23.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_orange_tokens_from_the_table_and_put_it_in_the_colourless_cup_Pick_up_the_blue_token_and_put_it_in_the_white_episode_0-train-bf_000240-af_000330-plus-0000938", "task": "Pick up the orange tokens from the table and put it in the colourless cup Pick up the blue token and put it in the white", "image": [ "images/droid_oxe/Pick_up_the_orange_tokens_from_the_table_and_put_it_in_the_colourless_cup_Pick_up_the_blue_token_and_put_it_in_the_white/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_orange_tokens_from_the_table_and_put_it_in_the_colourless_cup_Pick_up_the_blue_token_and_put_it_in_the_white/episode_0/cam_high/frame_000645.jpg", "images/droid_oxe/Pick_up_the_orange_tokens_from_the_table_and_put_it_in_the_colourless_cup_Pick_up_the_blue_token_and_put_it_in_the_white/episode_0/cam_high/frame_000240.jpg", "images/droid_oxe/Pick_up_the_orange_tokens_from_the_table_and_put_it_in_the_colourless_cup_Pick_up_the_blue_token_and_put_it_in_the_white/episode_0/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Pick_up_the_orange_tokens_from_the_table_and_put_it_in_the_colourless_cup_Pick_up_the_blue_token_and_put_it_in_the_white/episode_0/cam_right_wrist/frame_000240.jpg", "images/droid_oxe/Pick_up_the_orange_tokens_from_the_table_and_put_it_in_the_colourless_cup_Pick_up_the_blue_token_and_put_it_in_the_white/episode_0/cam_high/frame_000330.jpg", "images/droid_oxe/Pick_up_the_orange_tokens_from_the_table_and_put_it_in_the_colourless_cup_Pick_up_the_blue_token_and_put_it_in_the_white/episode_0/cam_left_wrist/frame_000330.jpg", "images/droid_oxe/Pick_up_the_orange_tokens_from_the_table_and_put_it_in_the_colourless_cup_Pick_up_the_blue_token_and_put_it_in_the_white/episode_0/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the orange tokens from the table and put it in the colourless cup Pick up the blue token and put it in the white`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the orange tokens from the table and put it in the colourless cup Pick up the blue token and put it in the white`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 22.2, "hop_absolute_value": 22.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_apple_to_the_left_take_the_radish_and_blocks_out_of_the_black_bowl_episode_0-train-bf_000315-af_000360-plus-0000914", "task": "Move the apple to the left take the radish and blocks out of the black bowl", "image": [ "images/droid_oxe/Move_the_apple_to_the_left_take_the_radish_and_blocks_out_of_the_black_bowl/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_apple_to_the_left_take_the_radish_and_blocks_out_of_the_black_bowl/episode_0/cam_high/frame_001055.jpg", "images/droid_oxe/Move_the_apple_to_the_left_take_the_radish_and_blocks_out_of_the_black_bowl/episode_0/cam_high/frame_000315.jpg", "images/droid_oxe/Move_the_apple_to_the_left_take_the_radish_and_blocks_out_of_the_black_bowl/episode_0/cam_left_wrist/frame_000315.jpg", "images/droid_oxe/Move_the_apple_to_the_left_take_the_radish_and_blocks_out_of_the_black_bowl/episode_0/cam_right_wrist/frame_000315.jpg", "images/droid_oxe/Move_the_apple_to_the_left_take_the_radish_and_blocks_out_of_the_black_bowl/episode_0/cam_high/frame_000360.jpg", "images/droid_oxe/Move_the_apple_to_the_left_take_the_radish_and_blocks_out_of_the_black_bowl/episode_0/cam_left_wrist/frame_000360.jpg", "images/droid_oxe/Move_the_apple_to_the_left_take_the_radish_and_blocks_out_of_the_black_bowl/episode_0/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the apple to the left take the radish and blocks out of the black bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the apple to the left take the radish and blocks out of the black bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 6.1, "hop_absolute_value": 6.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_blocks_in_the_cup_episode_2-train-bf_000015-af_000375-plus-0000493", "task": "Put the blocks in the cup", "image": [ "images/droid_oxe/Put_the_blocks_in_the_cup/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_blocks_in_the_cup/episode_1/cam_high/frame_000441.jpg", "images/droid_oxe/Put_the_blocks_in_the_cup/episode_2/cam_high/frame_000015.jpg", "images/droid_oxe/Put_the_blocks_in_the_cup/episode_2/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Put_the_blocks_in_the_cup/episode_2/cam_right_wrist/frame_000015.jpg", "images/droid_oxe/Put_the_blocks_in_the_cup/episode_2/cam_high/frame_000375.jpg", "images/droid_oxe/Put_the_blocks_in_the_cup/episode_2/cam_left_wrist/frame_000375.jpg", "images/droid_oxe/Put_the_blocks_in_the_cup/episode_2/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the blocks in the cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the blocks in the cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 92.3, "hop_absolute_value": 92.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_white_container_to_the_right_side_of_the_toaster_episode_0-train-bf_000000-af_000135-plus-0000226", "task": "Move the white container to the right side of the toaster", "image": [ "images/droid_oxe/Move_the_white_container_to_the_right_side_of_the_toaster/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_white_container_to_the_right_side_of_the_toaster/episode_0/cam_high/frame_000266.jpg", "images/droid_oxe/Move_the_white_container_to_the_right_side_of_the_toaster/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_white_container_to_the_right_side_of_the_toaster/episode_0/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Move_the_white_container_to_the_right_side_of_the_toaster/episode_0/cam_right_wrist/frame_000000.jpg", "images/droid_oxe/Move_the_white_container_to_the_right_side_of_the_toaster/episode_0/cam_high/frame_000135.jpg", "images/droid_oxe/Move_the_white_container_to_the_right_side_of_the_toaster/episode_0/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Move_the_white_container_to_the_right_side_of_the_toaster/episode_0/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the white container to the right side of the toaster`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the white container to the right side of the toaster`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 52.9, "hop_absolute_value": 52.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_three_bottles_to_the_left_one_at_a_time_episode_0-train-bf_000135-af_000270-plus-0000752", "task": "Move the three bottles to the left one at a time", "image": [ "images/droid_oxe/Move_the_three_bottles_to_the_left_one_at_a_time/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_three_bottles_to_the_left_one_at_a_time/episode_0/cam_high/frame_000435.jpg", "images/droid_oxe/Move_the_three_bottles_to_the_left_one_at_a_time/episode_0/cam_high/frame_000135.jpg", "images/droid_oxe/Move_the_three_bottles_to_the_left_one_at_a_time/episode_0/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Move_the_three_bottles_to_the_left_one_at_a_time/episode_0/cam_right_wrist/frame_000135.jpg", "images/droid_oxe/Move_the_three_bottles_to_the_left_one_at_a_time/episode_0/cam_high/frame_000270.jpg", "images/droid_oxe/Move_the_three_bottles_to_the_left_one_at_a_time/episode_0/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Move_the_three_bottles_to_the_left_one_at_a_time/episode_0/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the three bottles to the left one at a time`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the three bottles to the left one at a time`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 45.0, "hop_absolute_value": 45.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_blue_packet_of_chips_to_the_right_episode_0-train-bf_000030-af_000165-plus-0000333", "task": "Move the blue packet of chips to the right", "image": [ "images/droid_oxe/Move_the_blue_packet_of_chips_to_the_right/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_blue_packet_of_chips_to_the_right/episode_0/cam_high/frame_000507.jpg", "images/droid_oxe/Move_the_blue_packet_of_chips_to_the_right/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Move_the_blue_packet_of_chips_to_the_right/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Move_the_blue_packet_of_chips_to_the_right/episode_0/cam_right_wrist/frame_000030.jpg", "images/droid_oxe/Move_the_blue_packet_of_chips_to_the_right/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Move_the_blue_packet_of_chips_to_the_right/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Move_the_blue_packet_of_chips_to_the_right/episode_0/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the blue packet of chips to the right`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the blue packet of chips to the right`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 29.0, "hop_absolute_value": 29.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Unfold_the_white_short_then_move_the_black_cap_to_the_left_episode_0-train-bf_000105-af_000150-plus-0000613", "task": "Unfold the white short then move the black cap to the left", "image": [ "images/droid_oxe/Unfold_the_white_short_then_move_the_black_cap_to_the_left/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Unfold_the_white_short_then_move_the_black_cap_to_the_left/episode_0/cam_high/frame_000441.jpg", "images/droid_oxe/Unfold_the_white_short_then_move_the_black_cap_to_the_left/episode_0/cam_high/frame_000105.jpg", "images/droid_oxe/Unfold_the_white_short_then_move_the_black_cap_to_the_left/episode_0/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Unfold_the_white_short_then_move_the_black_cap_to_the_left/episode_0/cam_right_wrist/frame_000105.jpg", "images/droid_oxe/Unfold_the_white_short_then_move_the_black_cap_to_the_left/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Unfold_the_white_short_then_move_the_black_cap_to_the_left/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Unfold_the_white_short_then_move_the_black_cap_to_the_left/episode_0/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the white short then move the black cap to the left`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the white short then move the black cap to the left`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 13.6, "hop_absolute_value": 13.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_whiteboard_eraser_to_clean_the_whiteboard_episode_0-train-bf_000165-af_000525-plus-0000522", "task": "Use the whiteboard eraser to clean the whiteboard", "image": [ "images/droid_oxe/Use_the_whiteboard_eraser_to_clean_the_whiteboard/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_whiteboard_eraser_to_clean_the_whiteboard/episode_0/cam_high/frame_000592.jpg", "images/droid_oxe/Use_the_whiteboard_eraser_to_clean_the_whiteboard/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Use_the_whiteboard_eraser_to_clean_the_whiteboard/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Use_the_whiteboard_eraser_to_clean_the_whiteboard/episode_0/cam_right_wrist/frame_000165.jpg", "images/droid_oxe/Use_the_whiteboard_eraser_to_clean_the_whiteboard/episode_0/cam_high/frame_000525.jpg", "images/droid_oxe/Use_the_whiteboard_eraser_to_clean_the_whiteboard/episode_0/cam_left_wrist/frame_000525.jpg", "images/droid_oxe/Use_the_whiteboard_eraser_to_clean_the_whiteboard/episode_0/cam_right_wrist/frame_000525.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the whiteboard eraser to clean the whiteboard`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the whiteboard eraser to clean the whiteboard`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_cloth_in_the_box_episode_0-train-bf_000105-af_000135-plus-0000394", "task": "Put the cloth in the box", "image": [ "images/droid_oxe/Put_the_cloth_in_the_box/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_cloth_in_the_box/episode_0/cam_high/frame_000304.jpg", "images/droid_oxe/Put_the_cloth_in_the_box/episode_0/cam_high/frame_000105.jpg", "images/droid_oxe/Put_the_cloth_in_the_box/episode_0/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Put_the_cloth_in_the_box/episode_0/cam_right_wrist/frame_000105.jpg", "images/droid_oxe/Put_the_cloth_in_the_box/episode_0/cam_high/frame_000135.jpg", "images/droid_oxe/Put_the_cloth_in_the_box/episode_0/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Put_the_cloth_in_the_box/episode_0/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the cloth in the box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the cloth in the box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 15.4, "hop_absolute_value": 15.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_sponge_and_put_it_on_the_black_pan_use_the_sponge_to_scrub_the_pan_and_place_the_sponge_back_on_the_countert_episode_0-train-bf_000075-af_000615-plus-0000450", "task": "Pick up the sponge and put it on the black pan use the sponge to scrub the pan and place the sponge back on the countert", "image": [ "images/droid_oxe/Pick_up_the_sponge_and_put_it_on_the_black_pan_use_the_sponge_to_scrub_the_pan_and_place_the_sponge_back_on_the_countert/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_sponge_and_put_it_on_the_black_pan_use_the_sponge_to_scrub_the_pan_and_place_the_sponge_back_on_the_countert/episode_0/cam_high/frame_000899.jpg", "images/droid_oxe/Pick_up_the_sponge_and_put_it_on_the_black_pan_use_the_sponge_to_scrub_the_pan_and_place_the_sponge_back_on_the_countert/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Pick_up_the_sponge_and_put_it_on_the_black_pan_use_the_sponge_to_scrub_the_pan_and_place_the_sponge_back_on_the_countert/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Pick_up_the_sponge_and_put_it_on_the_black_pan_use_the_sponge_to_scrub_the_pan_and_place_the_sponge_back_on_the_countert/episode_0/cam_right_wrist/frame_000075.jpg", "images/droid_oxe/Pick_up_the_sponge_and_put_it_on_the_black_pan_use_the_sponge_to_scrub_the_pan_and_place_the_sponge_back_on_the_countert/episode_0/cam_high/frame_000615.jpg", "images/droid_oxe/Pick_up_the_sponge_and_put_it_on_the_black_pan_use_the_sponge_to_scrub_the_pan_and_place_the_sponge_back_on_the_countert/episode_0/cam_left_wrist/frame_000615.jpg", "images/droid_oxe/Pick_up_the_sponge_and_put_it_on_the_black_pan_use_the_sponge_to_scrub_the_pan_and_place_the_sponge_back_on_the_countert/episode_0/cam_right_wrist/frame_000615.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the sponge and put it on the black pan use the sponge to scrub the pan and place the sponge back on the countert`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the sponge and put it on the black pan use the sponge to scrub the pan and place the sponge back on the countert`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pour_the_contents_in_the_silver_bowl_into_the_blue_bowl_episode_0-train-bf_000270-af_000362-plus-0000425", "task": "Pour the contents in the silver bowl into the blue bowl", "image": [ "images/droid_oxe/Pour_the_contents_in_the_silver_bowl_into_the_blue_bowl/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pour_the_contents_in_the_silver_bowl_into_the_blue_bowl/episode_0/cam_high/frame_000362.jpg", "images/droid_oxe/Pour_the_contents_in_the_silver_bowl_into_the_blue_bowl/episode_0/cam_high/frame_000270.jpg", "images/droid_oxe/Pour_the_contents_in_the_silver_bowl_into_the_blue_bowl/episode_0/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Pour_the_contents_in_the_silver_bowl_into_the_blue_bowl/episode_0/cam_right_wrist/frame_000270.jpg", "images/droid_oxe/Pour_the_contents_in_the_silver_bowl_into_the_blue_bowl/episode_0/cam_high/frame_000362.jpg", "images/droid_oxe/Pour_the_contents_in_the_silver_bowl_into_the_blue_bowl/episode_0/cam_left_wrist/frame_000362.jpg", "images/droid_oxe/Pour_the_contents_in_the_silver_bowl_into_the_blue_bowl/episode_0/cam_right_wrist/frame_000362.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pour the contents in the silver bowl into the blue bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pour the contents in the silver bowl into the blue bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_yellow_towel_to_wipe_the_table_episode_8-train-bf_000330-af_000765-plus-0000501", "task": "Use the yellow towel to wipe the table", "image": [ "images/droid_oxe/Use_the_yellow_towel_to_wipe_the_table/episode_18/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_yellow_towel_to_wipe_the_table/episode_18/cam_high/frame_000202.jpg", "images/droid_oxe/Use_the_yellow_towel_to_wipe_the_table/episode_8/cam_high/frame_000330.jpg", "images/droid_oxe/Use_the_yellow_towel_to_wipe_the_table/episode_8/cam_left_wrist/frame_000330.jpg", "images/droid_oxe/Use_the_yellow_towel_to_wipe_the_table/episode_8/cam_right_wrist/frame_000330.jpg", "images/droid_oxe/Use_the_yellow_towel_to_wipe_the_table/episode_8/cam_high/frame_000765.jpg", "images/droid_oxe/Use_the_yellow_towel_to_wipe_the_table/episode_8/cam_left_wrist/frame_000765.jpg", "images/droid_oxe/Use_the_yellow_towel_to_wipe_the_table/episode_8/cam_right_wrist/frame_000765.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the yellow towel to wipe the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the yellow towel to wipe the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 76.3, "hop_absolute_value": 76.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Open_the_top_cabinet_door_and_put_an_orange_plushie_and_the_maize_in_the_top_compartment_of_the_cabinet_episode_0-train-bf_000225-af_000375-plus-0000943", "task": "Open the top cabinet door and put an orange plushie and the maize in the top compartment of the cabinet", "image": [ "images/droid_oxe/Open_the_top_cabinet_door_and_put_an_orange_plushie_and_the_maize_in_the_top_compartment_of_the_cabinet/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Open_the_top_cabinet_door_and_put_an_orange_plushie_and_the_maize_in_the_top_compartment_of_the_cabinet/episode_0/cam_high/frame_000505.jpg", "images/droid_oxe/Open_the_top_cabinet_door_and_put_an_orange_plushie_and_the_maize_in_the_top_compartment_of_the_cabinet/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Open_the_top_cabinet_door_and_put_an_orange_plushie_and_the_maize_in_the_top_compartment_of_the_cabinet/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Open_the_top_cabinet_door_and_put_an_orange_plushie_and_the_maize_in_the_top_compartment_of_the_cabinet/episode_0/cam_right_wrist/frame_000225.jpg", "images/droid_oxe/Open_the_top_cabinet_door_and_put_an_orange_plushie_and_the_maize_in_the_top_compartment_of_the_cabinet/episode_0/cam_high/frame_000375.jpg", "images/droid_oxe/Open_the_top_cabinet_door_and_put_an_orange_plushie_and_the_maize_in_the_top_compartment_of_the_cabinet/episode_0/cam_left_wrist/frame_000375.jpg", "images/droid_oxe/Open_the_top_cabinet_door_and_put_an_orange_plushie_and_the_maize_in_the_top_compartment_of_the_cabinet/episode_0/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the top cabinet door and put an orange plushie and the maize in the top compartment of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the top cabinet door and put an orange plushie and the maize in the top compartment of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 55.6, "hop_absolute_value": 55.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Push_the_novels_close_to_the_edge_move_the_sneaker_north_of_the_novels_and_then_use_the_napkin_to_wipe_the_sneaker_episode_0-train-bf_000060-af_000120-plus-0000624", "task": "Push the novels close to the edge move the sneaker north of the novels and then use the napkin to wipe the sneaker", "image": [ "images/droid_oxe/Push_the_novels_close_to_the_edge_move_the_sneaker_north_of_the_novels_and_then_use_the_napkin_to_wipe_the_sneaker/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Push_the_novels_close_to_the_edge_move_the_sneaker_north_of_the_novels_and_then_use_the_napkin_to_wipe_the_sneaker/episode_0/cam_high/frame_000447.jpg", "images/droid_oxe/Push_the_novels_close_to_the_edge_move_the_sneaker_north_of_the_novels_and_then_use_the_napkin_to_wipe_the_sneaker/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Push_the_novels_close_to_the_edge_move_the_sneaker_north_of_the_novels_and_then_use_the_napkin_to_wipe_the_sneaker/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Push_the_novels_close_to_the_edge_move_the_sneaker_north_of_the_novels_and_then_use_the_napkin_to_wipe_the_sneaker/episode_0/cam_right_wrist/frame_000060.jpg", "images/droid_oxe/Push_the_novels_close_to_the_edge_move_the_sneaker_north_of_the_novels_and_then_use_the_napkin_to_wipe_the_sneaker/episode_0/cam_high/frame_000120.jpg", "images/droid_oxe/Push_the_novels_close_to_the_edge_move_the_sneaker_north_of_the_novels_and_then_use_the_napkin_to_wipe_the_sneaker/episode_0/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Push_the_novels_close_to_the_edge_move_the_sneaker_north_of_the_novels_and_then_use_the_napkin_to_wipe_the_sneaker/episode_0/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the novels close to the edge move the sneaker north of the novels and then use the napkin to wipe the sneaker`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the novels close to the edge move the sneaker north of the novels and then use the napkin to wipe the sneaker`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.0, "hop_absolute_value": 16.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_measuring_cup_and_put_it_on_the_red_plate_episode_0-train-bf_000135-af_000150-plus-0000194", "task": "Pick up the measuring cup and put it on the red plate", "image": [ "images/droid_oxe/Pick_up_the_measuring_cup_and_put_it_on_the_red_plate/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_measuring_cup_and_put_it_on_the_red_plate/episode_0/cam_high/frame_000308.jpg", "images/droid_oxe/Pick_up_the_measuring_cup_and_put_it_on_the_red_plate/episode_0/cam_high/frame_000135.jpg", "images/droid_oxe/Pick_up_the_measuring_cup_and_put_it_on_the_red_plate/episode_0/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Pick_up_the_measuring_cup_and_put_it_on_the_red_plate/episode_0/cam_right_wrist/frame_000135.jpg", "images/droid_oxe/Pick_up_the_measuring_cup_and_put_it_on_the_red_plate/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Pick_up_the_measuring_cup_and_put_it_on_the_red_plate/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Pick_up_the_measuring_cup_and_put_it_on_the_red_plate/episode_0/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the measuring cup and put it on the red plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the measuring cup and put it on the red plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_the_bag_of_popcorn_and_pour_its_contents_on_the_cloth_on_the_table_episode_0-train-bf_000195-af_000990-plus-0000555", "task": "Pick the bag of popcorn and pour its contents on the cloth on the table", "image": [ "images/droid_oxe/Pick_the_bag_of_popcorn_and_pour_its_contents_on_the_cloth_on_the_table/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_the_bag_of_popcorn_and_pour_its_contents_on_the_cloth_on_the_table/episode_0/cam_high/frame_001148.jpg", "images/droid_oxe/Pick_the_bag_of_popcorn_and_pour_its_contents_on_the_cloth_on_the_table/episode_0/cam_high/frame_000195.jpg", "images/droid_oxe/Pick_the_bag_of_popcorn_and_pour_its_contents_on_the_cloth_on_the_table/episode_0/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Pick_the_bag_of_popcorn_and_pour_its_contents_on_the_cloth_on_the_table/episode_0/cam_right_wrist/frame_000195.jpg", "images/droid_oxe/Pick_the_bag_of_popcorn_and_pour_its_contents_on_the_cloth_on_the_table/episode_0/cam_high/frame_000990.jpg", "images/droid_oxe/Pick_the_bag_of_popcorn_and_pour_its_contents_on_the_cloth_on_the_table/episode_0/cam_left_wrist/frame_000990.jpg", "images/droid_oxe/Pick_the_bag_of_popcorn_and_pour_its_contents_on_the_cloth_on_the_table/episode_0/cam_right_wrist/frame_000990.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the bag of popcorn and pour its contents on the cloth on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the bag of popcorn and pour its contents on the cloth on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 84.1, "hop_absolute_value": 84.1, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_silver_bowl_and_blue_cup_to_the_washing_machine_on_the_right_episode_0-train-bf_000285-af_000300-plus-0000565", "task": "Move the silver bowl and blue cup to the washing machine on the right", "image": [ "images/droid_oxe/Move_the_silver_bowl_and_blue_cup_to_the_washing_machine_on_the_right/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_silver_bowl_and_blue_cup_to_the_washing_machine_on_the_right/episode_0/cam_high/frame_000460.jpg", "images/droid_oxe/Move_the_silver_bowl_and_blue_cup_to_the_washing_machine_on_the_right/episode_0/cam_high/frame_000285.jpg", "images/droid_oxe/Move_the_silver_bowl_and_blue_cup_to_the_washing_machine_on_the_right/episode_0/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/Move_the_silver_bowl_and_blue_cup_to_the_washing_machine_on_the_right/episode_0/cam_right_wrist/frame_000285.jpg", "images/droid_oxe/Move_the_silver_bowl_and_blue_cup_to_the_washing_machine_on_the_right/episode_0/cam_high/frame_000300.jpg", "images/droid_oxe/Move_the_silver_bowl_and_blue_cup_to_the_washing_machine_on_the_right/episode_0/cam_left_wrist/frame_000300.jpg", "images/droid_oxe/Move_the_silver_bowl_and_blue_cup_to_the_washing_machine_on_the_right/episode_0/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the silver bowl and blue cup to the washing machine on the right`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the silver bowl and blue cup to the washing machine on the right`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Wipe_the_top_part_of_the_dresser_with_a_tissue_episode_0-train-bf_000015-af_000360-plus-0000470", "task": "Wipe the top part of the dresser with a tissue", "image": [ "images/droid_oxe/Wipe_the_top_part_of_the_dresser_with_a_tissue/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Wipe_the_top_part_of_the_dresser_with_a_tissue/episode_0/cam_high/frame_000425.jpg", "images/droid_oxe/Wipe_the_top_part_of_the_dresser_with_a_tissue/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Wipe_the_top_part_of_the_dresser_with_a_tissue/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Wipe_the_top_part_of_the_dresser_with_a_tissue/episode_0/cam_right_wrist/frame_000015.jpg", "images/droid_oxe/Wipe_the_top_part_of_the_dresser_with_a_tissue/episode_0/cam_high/frame_000360.jpg", "images/droid_oxe/Wipe_the_top_part_of_the_dresser_with_a_tissue/episode_0/cam_left_wrist/frame_000360.jpg", "images/droid_oxe/Wipe_the_top_part_of_the_dresser_with_a_tissue/episode_0/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wipe the top part of the dresser with a tissue`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wipe the top part of the dresser with a tissue`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 85.2, "hop_absolute_value": 85.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_marker_in_the_white_mug_cup_episode_3-train-bf_000300-af_000320-plus-0000372", "task": "Put the marker in the white mug cup", "image": [ "images/droid_oxe/Put_the_marker_in_the_white_mug_cup/episode_2/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_marker_in_the_white_mug_cup/episode_2/cam_high/frame_000152.jpg", "images/droid_oxe/Put_the_marker_in_the_white_mug_cup/episode_3/cam_high/frame_000300.jpg", "images/droid_oxe/Put_the_marker_in_the_white_mug_cup/episode_3/cam_left_wrist/frame_000300.jpg", "images/droid_oxe/Put_the_marker_in_the_white_mug_cup/episode_3/cam_right_wrist/frame_000300.jpg", "images/droid_oxe/Put_the_marker_in_the_white_mug_cup/episode_3/cam_high/frame_000320.jpg", "images/droid_oxe/Put_the_marker_in_the_white_mug_cup/episode_3/cam_left_wrist/frame_000320.jpg", "images/droid_oxe/Put_the_marker_in_the_white_mug_cup/episode_3/cam_right_wrist/frame_000320.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the marker in the white mug cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the marker in the white mug cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Take_the_marker_out_of_the_pot_and_put_it_on_the_table_episode_0-train-bf_000000-af_000015-plus-0000091", "task": "Take the marker out of the pot and put it on the table", "image": [ "images/droid_oxe/Take_the_marker_out_of_the_pot_and_put_it_on_the_table/episode_3/cam_high/frame_000000.jpg", "images/droid_oxe/Take_the_marker_out_of_the_pot_and_put_it_on_the_table/episode_3/cam_high/frame_000153.jpg", "images/droid_oxe/Take_the_marker_out_of_the_pot_and_put_it_on_the_table/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Take_the_marker_out_of_the_pot_and_put_it_on_the_table/episode_0/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Take_the_marker_out_of_the_pot_and_put_it_on_the_table/episode_0/cam_right_wrist/frame_000000.jpg", "images/droid_oxe/Take_the_marker_out_of_the_pot_and_put_it_on_the_table/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Take_the_marker_out_of_the_pot_and_put_it_on_the_table/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Take_the_marker_out_of_the_pot_and_put_it_on_the_table/episode_0/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the marker out of the pot and put it on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the marker out of the pot and put it on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 7.7, "hop_absolute_value": 7.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_cups_in_the_dishes_tray_upside_down_episode_0-train-bf_000015-af_000240-plus-0000362", "task": "Put the cups in the dishes tray upside down", "image": [ "images/droid_oxe/Put_the_cups_in_the_dishes_tray_upside_down/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_cups_in_the_dishes_tray_upside_down/episode_0/cam_high/frame_000464.jpg", "images/droid_oxe/Put_the_cups_in_the_dishes_tray_upside_down/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Put_the_cups_in_the_dishes_tray_upside_down/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Put_the_cups_in_the_dishes_tray_upside_down/episode_0/cam_right_wrist/frame_000015.jpg", "images/droid_oxe/Put_the_cups_in_the_dishes_tray_upside_down/episode_0/cam_high/frame_000240.jpg", "images/droid_oxe/Put_the_cups_in_the_dishes_tray_upside_down/episode_0/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Put_the_cups_in_the_dishes_tray_upside_down/episode_0/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the cups in the dishes tray upside down`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the cups in the dishes tray upside down`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 51.7, "hop_absolute_value": 51.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_silver_lid_to_the_counter_and_then_place_it_back_on_the_clear_object_episode_0-train-bf_000060-af_000180-plus-0000639", "task": "Move the silver lid to the counter and then place it back on the clear object", "image": [ "images/droid_oxe/Move_the_silver_lid_to_the_counter_and_then_place_it_back_on_the_clear_object/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_silver_lid_to_the_counter_and_then_place_it_back_on_the_clear_object/episode_0/cam_high/frame_000396.jpg", "images/droid_oxe/Move_the_silver_lid_to_the_counter_and_then_place_it_back_on_the_clear_object/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Move_the_silver_lid_to_the_counter_and_then_place_it_back_on_the_clear_object/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Move_the_silver_lid_to_the_counter_and_then_place_it_back_on_the_clear_object/episode_0/cam_right_wrist/frame_000060.jpg", "images/droid_oxe/Move_the_silver_lid_to_the_counter_and_then_place_it_back_on_the_clear_object/episode_0/cam_high/frame_000180.jpg", "images/droid_oxe/Move_the_silver_lid_to_the_counter_and_then_place_it_back_on_the_clear_object/episode_0/cam_left_wrist/frame_000180.jpg", "images/droid_oxe/Move_the_silver_lid_to_the_counter_and_then_place_it_back_on_the_clear_object/episode_0/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the silver lid to the counter and then place it back on the clear object`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the silver lid to the counter and then place it back on the clear object`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 36.4, "hop_absolute_value": 36.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_pot_to_the_front_plate_episode_0-train-bf_000075-af_000090-plus-0000050", "task": "Move the pot to the front plate", "image": [ "images/droid_oxe/Move_the_pot_to_the_front_plate/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_pot_to_the_front_plate/episode_0/cam_high/frame_000141.jpg", "images/droid_oxe/Move_the_pot_to_the_front_plate/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Move_the_pot_to_the_front_plate/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Move_the_pot_to_the_front_plate/episode_0/cam_right_wrist/frame_000075.jpg", "images/droid_oxe/Move_the_pot_to_the_front_plate/episode_0/cam_high/frame_000090.jpg", "images/droid_oxe/Move_the_pot_to_the_front_plate/episode_0/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Move_the_pot_to_the_front_plate/episode_0/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the pot to the front plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the pot to the front plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_blue_fabric_to_wipe_the_stove_episode_0-train-bf_000015-af_000030-plus-0000431", "task": "Use the blue fabric to wipe the stove", "image": [ "images/droid_oxe/Use_the_blue_fabric_to_wipe_the_stove/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_blue_fabric_to_wipe_the_stove/episode_0/cam_high/frame_000344.jpg", "images/droid_oxe/Use_the_blue_fabric_to_wipe_the_stove/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Use_the_blue_fabric_to_wipe_the_stove/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Use_the_blue_fabric_to_wipe_the_stove/episode_0/cam_right_wrist/frame_000015.jpg", "images/droid_oxe/Use_the_blue_fabric_to_wipe_the_stove/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Use_the_blue_fabric_to_wipe_the_stove/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Use_the_blue_fabric_to_wipe_the_stove/episode_0/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the blue fabric to wipe the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the blue fabric to wipe the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.8, "hop_absolute_value": 4.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Separate_the_three_pots_episode_0-train-bf_000060-af_000090-plus-0000474", "task": "Separate the three pots", "image": [ "images/droid_oxe/Separate_the_three_pots/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Separate_the_three_pots/episode_0/cam_high/frame_000366.jpg", "images/droid_oxe/Separate_the_three_pots/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Separate_the_three_pots/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Separate_the_three_pots/episode_0/cam_right_wrist/frame_000060.jpg", "images/droid_oxe/Separate_the_three_pots/episode_0/cam_high/frame_000090.jpg", "images/droid_oxe/Separate_the_three_pots/episode_0/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Separate_the_three_pots/episode_0/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Separate the three pots`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Separate the three pots`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Clean_the_table_episode_3-train-bf_000090-af_000285-plus-0000545", "task": "Clean the table", "image": [ "images/droid_oxe/Clean_the_table/episode_3/cam_high/frame_000000.jpg", "images/droid_oxe/Clean_the_table/episode_3/cam_high/frame_000352.jpg", "images/droid_oxe/Clean_the_table/episode_3/cam_high/frame_000090.jpg", "images/droid_oxe/Clean_the_table/episode_3/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Clean_the_table/episode_3/cam_right_wrist/frame_000090.jpg", "images/droid_oxe/Clean_the_table/episode_3/cam_high/frame_000285.jpg", "images/droid_oxe/Clean_the_table/episode_3/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/Clean_the_table/episode_3/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Clean the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Clean the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 76.5, "hop_absolute_value": 76.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_fleece_from_the_backrest_of_the_chair_episode_0-train-bf_000195-af_000210-plus-0000174", "task": "Remove the fleece from the backrest of the chair", "image": [ "images/droid_oxe/Remove_the_fleece_from_the_backrest_of_the_chair/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_fleece_from_the_backrest_of_the_chair/episode_0/cam_high/frame_000262.jpg", "images/droid_oxe/Remove_the_fleece_from_the_backrest_of_the_chair/episode_0/cam_high/frame_000195.jpg", "images/droid_oxe/Remove_the_fleece_from_the_backrest_of_the_chair/episode_0/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Remove_the_fleece_from_the_backrest_of_the_chair/episode_0/cam_right_wrist/frame_000195.jpg", "images/droid_oxe/Remove_the_fleece_from_the_backrest_of_the_chair/episode_0/cam_high/frame_000210.jpg", "images/droid_oxe/Remove_the_fleece_from_the_backrest_of_the_chair/episode_0/cam_left_wrist/frame_000210.jpg", "images/droid_oxe/Remove_the_fleece_from_the_backrest_of_the_chair/episode_0/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the fleece from the backrest of the chair`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the fleece from the backrest of the chair`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_rod_and_move_it_within_the_silver_cup_episode_0-train-bf_000060-af_000075-plus-0000109", "task": "Pick up the rod and move it within the silver cup", "image": [ "images/droid_oxe/Pick_up_the_rod_and_move_it_within_the_silver_cup/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_rod_and_move_it_within_the_silver_cup/episode_0/cam_high/frame_000215.jpg", "images/droid_oxe/Pick_up_the_rod_and_move_it_within_the_silver_cup/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Pick_up_the_rod_and_move_it_within_the_silver_cup/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Pick_up_the_rod_and_move_it_within_the_silver_cup/episode_0/cam_right_wrist/frame_000060.jpg", "images/droid_oxe/Pick_up_the_rod_and_move_it_within_the_silver_cup/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Pick_up_the_rod_and_move_it_within_the_silver_cup/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Pick_up_the_rod_and_move_it_within_the_silver_cup/episode_0/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the rod and move it within the silver cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the rod and move it within the silver cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_spoon_and_stir_the_objects_in_the_bowl_episode_0-train-bf_000240-af_000345-plus-0000349", "task": "Pick up the spoon and stir the objects in the bowl", "image": [ "images/droid_oxe/Pick_up_the_spoon_and_stir_the_objects_in_the_bowl/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_spoon_and_stir_the_objects_in_the_bowl/episode_0/cam_high/frame_000468.jpg", "images/droid_oxe/Pick_up_the_spoon_and_stir_the_objects_in_the_bowl/episode_0/cam_high/frame_000240.jpg", "images/droid_oxe/Pick_up_the_spoon_and_stir_the_objects_in_the_bowl/episode_0/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Pick_up_the_spoon_and_stir_the_objects_in_the_bowl/episode_0/cam_right_wrist/frame_000240.jpg", "images/droid_oxe/Pick_up_the_spoon_and_stir_the_objects_in_the_bowl/episode_0/cam_high/frame_000345.jpg", "images/droid_oxe/Pick_up_the_spoon_and_stir_the_objects_in_the_bowl/episode_0/cam_left_wrist/frame_000345.jpg", "images/droid_oxe/Pick_up_the_spoon_and_stir_the_objects_in_the_bowl/episode_0/cam_right_wrist/frame_000345.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the spoon and stir the objects in the bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the spoon and stir the objects in the bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 46.7, "hop_absolute_value": 46.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_string_from_the_table_and_make_a_knot_episode_0-train-bf_000345-af_000360-plus-0000623", "task": "Use the string from the table and make a knot", "image": [ "images/droid_oxe/Use_the_string_from_the_table_and_make_a_knot/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_string_from_the_table_and_make_a_knot/episode_0/cam_high/frame_000889.jpg", "images/droid_oxe/Use_the_string_from_the_table_and_make_a_knot/episode_0/cam_high/frame_000345.jpg", "images/droid_oxe/Use_the_string_from_the_table_and_make_a_knot/episode_0/cam_left_wrist/frame_000345.jpg", "images/droid_oxe/Use_the_string_from_the_table_and_make_a_knot/episode_0/cam_right_wrist/frame_000345.jpg", "images/droid_oxe/Use_the_string_from_the_table_and_make_a_knot/episode_0/cam_high/frame_000360.jpg", "images/droid_oxe/Use_the_string_from_the_table_and_make_a_knot/episode_0/cam_left_wrist/frame_000360.jpg", "images/droid_oxe/Use_the_string_from_the_table_and_make_a_knot/episode_0/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the string from the table and make a knot`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the string from the table and make a knot`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.8, "hop_absolute_value": 2.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_silver_lid_on_the_silver_bottle_episode_0-train-bf_000135-af_000720-plus-0000588", "task": "Put the silver lid on the silver bottle", "image": [ "images/droid_oxe/Put_the_silver_lid_on_the_silver_bottle/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_silver_lid_on_the_silver_bottle/episode_0/cam_high/frame_000755.jpg", "images/droid_oxe/Put_the_silver_lid_on_the_silver_bottle/episode_0/cam_high/frame_000135.jpg", "images/droid_oxe/Put_the_silver_lid_on_the_silver_bottle/episode_0/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Put_the_silver_lid_on_the_silver_bottle/episode_0/cam_right_wrist/frame_000135.jpg", "images/droid_oxe/Put_the_silver_lid_on_the_silver_bottle/episode_0/cam_high/frame_000720.jpg", "images/droid_oxe/Put_the_silver_lid_on_the_silver_bottle/episode_0/cam_left_wrist/frame_000720.jpg", "images/droid_oxe/Put_the_silver_lid_on_the_silver_bottle/episode_0/cam_right_wrist/frame_000720.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the silver lid on the silver bottle`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the silver lid on the silver bottle`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 95.1, "hop_absolute_value": 95.1, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_jar_slightly_to_the_left_episode_0-train-bf_000180-af_000195-plus-0000198", "task": "Move the jar slightly to the left", "image": [ "images/droid_oxe/Move_the_jar_slightly_to_the_left/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_jar_slightly_to_the_left/episode_0/cam_high/frame_000314.jpg", "images/droid_oxe/Move_the_jar_slightly_to_the_left/episode_0/cam_high/frame_000180.jpg", "images/droid_oxe/Move_the_jar_slightly_to_the_left/episode_0/cam_left_wrist/frame_000180.jpg", "images/droid_oxe/Move_the_jar_slightly_to_the_left/episode_0/cam_right_wrist/frame_000180.jpg", "images/droid_oxe/Move_the_jar_slightly_to_the_left/episode_0/cam_high/frame_000195.jpg", "images/droid_oxe/Move_the_jar_slightly_to_the_left/episode_0/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Move_the_jar_slightly_to_the_left/episode_0/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the jar slightly to the left`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the jar slightly to the left`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_rolling_pin_to_the_left_and_right_twice_episode_0-train-bf_000075-af_000165-plus-0000186", "task": "Move the rolling pin to the left and right twice", "image": [ "images/droid_oxe/Move_the_rolling_pin_to_the_left_and_right_twice/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_rolling_pin_to_the_left_and_right_twice/episode_0/cam_high/frame_000226.jpg", "images/droid_oxe/Move_the_rolling_pin_to_the_left_and_right_twice/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Move_the_rolling_pin_to_the_left_and_right_twice/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Move_the_rolling_pin_to_the_left_and_right_twice/episode_0/cam_right_wrist/frame_000075.jpg", "images/droid_oxe/Move_the_rolling_pin_to_the_left_and_right_twice/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Move_the_rolling_pin_to_the_left_and_right_twice/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Move_the_rolling_pin_to_the_left_and_right_twice/episode_0/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the rolling pin to the left and right twice`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the rolling pin to the left and right twice`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Flip_the_green_bowl_over_episode_0-train-bf_000060-af_000150-plus-0000128", "task": "Flip the green bowl over", "image": [ "images/droid_oxe/Flip_the_green_bowl_over/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Flip_the_green_bowl_over/episode_0/cam_high/frame_000193.jpg", "images/droid_oxe/Flip_the_green_bowl_over/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Flip_the_green_bowl_over/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Flip_the_green_bowl_over/episode_0/cam_right_wrist/frame_000060.jpg", "images/droid_oxe/Flip_the_green_bowl_over/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Flip_the_green_bowl_over/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Flip_the_green_bowl_over/episode_0/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Flip the green bowl over`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Flip the green bowl over`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_wooden_spoon_from_the_pot_on_the_right_side_of_the_stove_and_place_it_on_the_left_side_of_the_countertop_episode_0-train-bf_000150-af_000165-plus-0000204", "task": "Remove the wooden spoon from the pot on the right side of the stove and place it on the left side of the countertop", "image": [ "images/droid_oxe/Remove_the_wooden_spoon_from_the_pot_on_the_right_side_of_the_stove_and_place_it_on_the_left_side_of_the_countertop/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_wooden_spoon_from_the_pot_on_the_right_side_of_the_stove_and_place_it_on_the_left_side_of_the_countertop/episode_0/cam_high/frame_000323.jpg", "images/droid_oxe/Remove_the_wooden_spoon_from_the_pot_on_the_right_side_of_the_stove_and_place_it_on_the_left_side_of_the_countertop/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Remove_the_wooden_spoon_from_the_pot_on_the_right_side_of_the_stove_and_place_it_on_the_left_side_of_the_countertop/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Remove_the_wooden_spoon_from_the_pot_on_the_right_side_of_the_stove_and_place_it_on_the_left_side_of_the_countertop/episode_0/cam_right_wrist/frame_000150.jpg", "images/droid_oxe/Remove_the_wooden_spoon_from_the_pot_on_the_right_side_of_the_stove_and_place_it_on_the_left_side_of_the_countertop/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Remove_the_wooden_spoon_from_the_pot_on_the_right_side_of_the_stove_and_place_it_on_the_left_side_of_the_countertop/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Remove_the_wooden_spoon_from_the_pot_on_the_right_side_of_the_stove_and_place_it_on_the_left_side_of_the_countertop/episode_0/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the wooden spoon from the pot on the right side of the stove and place it on the left side of the countertop`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the wooden spoon from the pot on the right side of the stove and place it on the left side of the countertop`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Empty_the_contents_in_the_clear_jar_episode_0-train-bf_000060-af_000210-plus-0000332", "task": "Empty the contents in the clear jar", "image": [ "images/droid_oxe/Empty_the_contents_in_the_clear_jar/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Empty_the_contents_in_the_clear_jar/episode_0/cam_high/frame_000390.jpg", "images/droid_oxe/Empty_the_contents_in_the_clear_jar/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Empty_the_contents_in_the_clear_jar/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Empty_the_contents_in_the_clear_jar/episode_0/cam_right_wrist/frame_000060.jpg", "images/droid_oxe/Empty_the_contents_in_the_clear_jar/episode_0/cam_high/frame_000210.jpg", "images/droid_oxe/Empty_the_contents_in_the_clear_jar/episode_0/cam_left_wrist/frame_000210.jpg", "images/droid_oxe/Empty_the_contents_in_the_clear_jar/episode_0/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Empty the contents in the clear jar`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Empty the contents in the clear jar`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 45.5, "hop_absolute_value": 45.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_faucet_of_the_sink_to_the_left_episode_0-train-bf_000045-af_000060-plus-0000080", "task": "Move the faucet of the sink to the left", "image": [ "images/droid_oxe/Move_the_faucet_of_the_sink_to_the_left/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_faucet_of_the_sink_to_the_left/episode_0/cam_high/frame_000193.jpg", "images/droid_oxe/Move_the_faucet_of_the_sink_to_the_left/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Move_the_faucet_of_the_sink_to_the_left/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Move_the_faucet_of_the_sink_to_the_left/episode_0/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Move_the_faucet_of_the_sink_to_the_left/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Move_the_faucet_of_the_sink_to_the_left/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Move_the_faucet_of_the_sink_to_the_left/episode_0/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the faucet of the sink to the left`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the faucet of the sink to the left`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_mug_forward_episode_19-train-bf_000060-af_000120-plus-0000084", "task": "Move the mug forward", "image": [ "images/droid_oxe/Move_the_mug_forward/episode_8/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_mug_forward/episode_8/cam_high/frame_000156.jpg", "images/droid_oxe/Move_the_mug_forward/episode_19/cam_high/frame_000060.jpg", "images/droid_oxe/Move_the_mug_forward/episode_19/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Move_the_mug_forward/episode_19/cam_right_wrist/frame_000060.jpg", "images/droid_oxe/Move_the_mug_forward/episode_19/cam_high/frame_000120.jpg", "images/droid_oxe/Move_the_mug_forward/episode_19/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Move_the_mug_forward/episode_19/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the mug forward`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the mug forward`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_plate_from_the_left_side_of_the_dishwasher_to_the_right_side_episode_0-train-bf_000045-af_000165-plus-0000317", "task": "Move the plate from the left side of the dishwasher to the right side", "image": [ "images/droid_oxe/Move_the_plate_from_the_left_side_of_the_dishwasher_to_the_right_side/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_plate_from_the_left_side_of_the_dishwasher_to_the_right_side/episode_0/cam_high/frame_000477.jpg", "images/droid_oxe/Move_the_plate_from_the_left_side_of_the_dishwasher_to_the_right_side/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Move_the_plate_from_the_left_side_of_the_dishwasher_to_the_right_side/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Move_the_plate_from_the_left_side_of_the_dishwasher_to_the_right_side/episode_0/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Move_the_plate_from_the_left_side_of_the_dishwasher_to_the_right_side/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Move_the_plate_from_the_left_side_of_the_dishwasher_to_the_right_side/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Move_the_plate_from_the_left_side_of_the_dishwasher_to_the_right_side/episode_0/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the plate from the left side of the dishwasher to the right side`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the plate from the left side of the dishwasher to the right side`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 28.6, "hop_absolute_value": 28.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Close_the_mouth_of_the_flask_episode_1-train-bf_000015-af_000105-plus-0000062", "task": "Close the mouth of the flask", "image": [ "images/droid_oxe/Close_the_mouth_of_the_flask/episode_3/cam_high/frame_000000.jpg", "images/droid_oxe/Close_the_mouth_of_the_flask/episode_3/cam_high/frame_000120.jpg", "images/droid_oxe/Close_the_mouth_of_the_flask/episode_1/cam_high/frame_000015.jpg", "images/droid_oxe/Close_the_mouth_of_the_flask/episode_1/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Close_the_mouth_of_the_flask/episode_1/cam_right_wrist/frame_000015.jpg", "images/droid_oxe/Close_the_mouth_of_the_flask/episode_1/cam_high/frame_000105.jpg", "images/droid_oxe/Close_the_mouth_of_the_flask/episode_1/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Close_the_mouth_of_the_flask/episode_1/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the mouth of the flask`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the mouth of the flask`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_green_can_backwards_episode_0-train-bf_000045-af_000165-plus-0000156", "task": "Move the green can backwards", "image": [ "images/droid_oxe/Move_the_green_can_backwards/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_green_can_backwards/episode_0/cam_high/frame_000198.jpg", "images/droid_oxe/Move_the_green_can_backwards/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Move_the_green_can_backwards/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Move_the_green_can_backwards/episode_0/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Move_the_green_can_backwards/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Move_the_green_can_backwards/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Move_the_green_can_backwards/episode_0/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the green can backwards`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the green can backwards`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_cup_from_the_dish_dryer_and_put_it_on_the_window_episode_0-train-bf_000240-af_000270-plus-0000518", "task": "Pick up the cup from the dish dryer and put it on the window", "image": [ "images/droid_oxe/Pick_up_the_cup_from_the_dish_dryer_and_put_it_on_the_window/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_cup_from_the_dish_dryer_and_put_it_on_the_window/episode_0/cam_high/frame_000350.jpg", "images/droid_oxe/Pick_up_the_cup_from_the_dish_dryer_and_put_it_on_the_window/episode_0/cam_high/frame_000240.jpg", "images/droid_oxe/Pick_up_the_cup_from_the_dish_dryer_and_put_it_on_the_window/episode_0/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Pick_up_the_cup_from_the_dish_dryer_and_put_it_on_the_window/episode_0/cam_right_wrist/frame_000240.jpg", "images/droid_oxe/Pick_up_the_cup_from_the_dish_dryer_and_put_it_on_the_window/episode_0/cam_high/frame_000270.jpg", "images/droid_oxe/Pick_up_the_cup_from_the_dish_dryer_and_put_it_on_the_window/episode_0/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Pick_up_the_cup_from_the_dish_dryer_and_put_it_on_the_window/episode_0/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the cup from the dish dryer and put it on the window`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the cup from the dish dryer and put it on the window`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 28.6, "hop_absolute_value": 28.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_white_towel_and_hang_it_next_to_the_other_towel_episode_0-train-bf_000030-af_000360-plus-0000465", "task": "Pick up the white towel and hang it next to the other towel", "image": [ "images/droid_oxe/Pick_up_the_white_towel_and_hang_it_next_to_the_other_towel/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_white_towel_and_hang_it_next_to_the_other_towel/episode_0/cam_high/frame_000407.jpg", "images/droid_oxe/Pick_up_the_white_towel_and_hang_it_next_to_the_other_towel/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Pick_up_the_white_towel_and_hang_it_next_to_the_other_towel/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Pick_up_the_white_towel_and_hang_it_next_to_the_other_towel/episode_0/cam_right_wrist/frame_000030.jpg", "images/droid_oxe/Pick_up_the_white_towel_and_hang_it_next_to_the_other_towel/episode_0/cam_high/frame_000360.jpg", "images/droid_oxe/Pick_up_the_white_towel_and_hang_it_next_to_the_other_towel/episode_0/cam_left_wrist/frame_000360.jpg", "images/droid_oxe/Pick_up_the_white_towel_and_hang_it_next_to_the_other_towel/episode_0/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the white towel and hang it next to the other towel`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the white towel and hang it next to the other towel`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 88.0, "hop_absolute_value": 88.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_spoon_to_transfer_one_spoon_of_contents_from_the_right_bowl_to_the_left_bowl_then_press_the_button_on_the_scale_episode_0-train-bf_000705-af_000720-plus-0000727", "task": "Use the spoon to transfer one spoon of contents from the right bowl to the left bowl then press the button on the scale", "image": [ "images/droid_oxe/Use_the_spoon_to_transfer_one_spoon_of_contents_from_the_right_bowl_to_the_left_bowl_then_press_the_button_on_the_scale/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_spoon_to_transfer_one_spoon_of_contents_from_the_right_bowl_to_the_left_bowl_then_press_the_button_on_the_scale/episode_0/cam_high/frame_001204.jpg", "images/droid_oxe/Use_the_spoon_to_transfer_one_spoon_of_contents_from_the_right_bowl_to_the_left_bowl_then_press_the_button_on_the_scale/episode_0/cam_high/frame_000705.jpg", "images/droid_oxe/Use_the_spoon_to_transfer_one_spoon_of_contents_from_the_right_bowl_to_the_left_bowl_then_press_the_button_on_the_scale/episode_0/cam_left_wrist/frame_000705.jpg", "images/droid_oxe/Use_the_spoon_to_transfer_one_spoon_of_contents_from_the_right_bowl_to_the_left_bowl_then_press_the_button_on_the_scale/episode_0/cam_right_wrist/frame_000705.jpg", "images/droid_oxe/Use_the_spoon_to_transfer_one_spoon_of_contents_from_the_right_bowl_to_the_left_bowl_then_press_the_button_on_the_scale/episode_0/cam_high/frame_000720.jpg", "images/droid_oxe/Use_the_spoon_to_transfer_one_spoon_of_contents_from_the_right_bowl_to_the_left_bowl_then_press_the_button_on_the_scale/episode_0/cam_left_wrist/frame_000720.jpg", "images/droid_oxe/Use_the_spoon_to_transfer_one_spoon_of_contents_from_the_right_bowl_to_the_left_bowl_then_press_the_button_on_the_scale/episode_0/cam_right_wrist/frame_000720.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the spoon to transfer one spoon of contents from the right bowl to the left bowl then press the button on the scale`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the spoon to transfer one spoon of contents from the right bowl to the left bowl then press the button on the scale`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.0, "hop_absolute_value": 3.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_spoon_and_scoop_some_of_the_contents_from_the_silver_bowl_and_put_it_in_the_other_silver_bowl_episode_0-train-bf_000030-af_000225-plus-0000343", "task": "Pick up the spoon and scoop some of the contents from the silver bowl and put it in the other silver bowl", "image": [ "images/droid_oxe/Pick_up_the_spoon_and_scoop_some_of_the_contents_from_the_silver_bowl_and_put_it_in_the_other_silver_bowl/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_spoon_and_scoop_some_of_the_contents_from_the_silver_bowl_and_put_it_in_the_other_silver_bowl/episode_0/cam_high/frame_000356.jpg", "images/droid_oxe/Pick_up_the_spoon_and_scoop_some_of_the_contents_from_the_silver_bowl_and_put_it_in_the_other_silver_bowl/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Pick_up_the_spoon_and_scoop_some_of_the_contents_from_the_silver_bowl_and_put_it_in_the_other_silver_bowl/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Pick_up_the_spoon_and_scoop_some_of_the_contents_from_the_silver_bowl_and_put_it_in_the_other_silver_bowl/episode_0/cam_right_wrist/frame_000030.jpg", "images/droid_oxe/Pick_up_the_spoon_and_scoop_some_of_the_contents_from_the_silver_bowl_and_put_it_in_the_other_silver_bowl/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Pick_up_the_spoon_and_scoop_some_of_the_contents_from_the_silver_bowl_and_put_it_in_the_other_silver_bowl/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Pick_up_the_spoon_and_scoop_some_of_the_contents_from_the_silver_bowl_and_put_it_in_the_other_silver_bowl/episode_0/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the spoon and scoop some of the contents from the silver bowl and put it in the other silver bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the spoon and scoop some of the contents from the silver bowl and put it in the other silver bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 61.9, "hop_absolute_value": 61.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Take_a_packet_out_of_the_box_and_place_it_on_the_table_episode_0-train-bf_000030-af_000135-plus-0000288", "task": "Take a packet out of the box and place it on the table", "image": [ "images/droid_oxe/Take_a_packet_out_of_the_box_and_place_it_on_the_table/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Take_a_packet_out_of_the_box_and_place_it_on_the_table/episode_0/cam_high/frame_000376.jpg", "images/droid_oxe/Take_a_packet_out_of_the_box_and_place_it_on_the_table/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Take_a_packet_out_of_the_box_and_place_it_on_the_table/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Take_a_packet_out_of_the_box_and_place_it_on_the_table/episode_0/cam_right_wrist/frame_000030.jpg", "images/droid_oxe/Take_a_packet_out_of_the_box_and_place_it_on_the_table/episode_0/cam_high/frame_000135.jpg", "images/droid_oxe/Take_a_packet_out_of_the_box_and_place_it_on_the_table/episode_0/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Take_a_packet_out_of_the_box_and_place_it_on_the_table/episode_0/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take a packet out of the box and place it on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take a packet out of the box and place it on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 30.4, "hop_absolute_value": 30.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Press_some_of_the_keys_on_the_keyboard_episode_0-train-bf_000045-af_000075-plus-0000132", "task": "Press some of the keys on the keyboard", "image": [ "images/droid_oxe/Press_some_of_the_keys_on_the_keyboard/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Press_some_of_the_keys_on_the_keyboard/episode_0/cam_high/frame_000230.jpg", "images/droid_oxe/Press_some_of_the_keys_on_the_keyboard/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Press_some_of_the_keys_on_the_keyboard/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Press_some_of_the_keys_on_the_keyboard/episode_0/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Press_some_of_the_keys_on_the_keyboard/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Press_some_of_the_keys_on_the_keyboard/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Press_some_of_the_keys_on_the_keyboard/episode_0/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Press some of the keys on the keyboard`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Press some of the keys on the keyboard`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_green_object_from_the_bowl_episode_0-train-bf_000120-af_000255-plus-0000295", "task": "Remove the green object from the bowl", "image": [ "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_1/cam_high/frame_000182.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_0/cam_high/frame_000120.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_0/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_0/cam_right_wrist/frame_000120.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_0/cam_high/frame_000255.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_0/cam_left_wrist/frame_000255.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_0/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the green object from the bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the green object from the bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 81.8, "hop_absolute_value": 81.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_towel_from_the_countertop_to_the_hanger_on_the_drawer_episode_0-train-bf_000045-af_000120-plus-0000210", "task": "Move the towel from the countertop to the hanger on the drawer", "image": [ "images/droid_oxe/Move_the_towel_from_the_countertop_to_the_hanger_on_the_drawer/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_towel_from_the_countertop_to_the_hanger_on_the_drawer/episode_0/cam_high/frame_000296.jpg", "images/droid_oxe/Move_the_towel_from_the_countertop_to_the_hanger_on_the_drawer/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Move_the_towel_from_the_countertop_to_the_hanger_on_the_drawer/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Move_the_towel_from_the_countertop_to_the_hanger_on_the_drawer/episode_0/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Move_the_towel_from_the_countertop_to_the_hanger_on_the_drawer/episode_0/cam_high/frame_000120.jpg", "images/droid_oxe/Move_the_towel_from_the_countertop_to_the_hanger_on_the_drawer/episode_0/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Move_the_towel_from_the_countertop_to_the_hanger_on_the_drawer/episode_0/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the towel from the countertop to the hanger on the drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the towel from the countertop to the hanger on the drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 31.2, "hop_absolute_value": 31.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_green_object_to_the_right_episode_4-train-bf_000015-af_000240-plus-0000828", "task": "Move the green object to the right", "image": [ "images/droid_oxe/Move_the_green_object_to_the_right/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_green_object_to_the_right/episode_0/cam_high/frame_000104.jpg", "images/droid_oxe/Move_the_green_object_to_the_right/episode_4/cam_high/frame_000015.jpg", "images/droid_oxe/Move_the_green_object_to_the_right/episode_4/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Move_the_green_object_to_the_right/episode_4/cam_right_wrist/frame_000015.jpg", "images/droid_oxe/Move_the_green_object_to_the_right/episode_4/cam_high/frame_000240.jpg", "images/droid_oxe/Move_the_green_object_to_the_right/episode_4/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Move_the_green_object_to_the_right/episode_4/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the green object to the right`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the green object to the right`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_bowl_on_the_table_episode_1-train-bf_000060-af_000105-plus-0000074", "task": "Move the bowl on the table", "image": [ "images/droid_oxe/Move_the_bowl_on_the_table/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_bowl_on_the_table/episode_1/cam_high/frame_000156.jpg", "images/droid_oxe/Move_the_bowl_on_the_table/episode_1/cam_high/frame_000060.jpg", "images/droid_oxe/Move_the_bowl_on_the_table/episode_1/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Move_the_bowl_on_the_table/episode_1/cam_right_wrist/frame_000060.jpg", "images/droid_oxe/Move_the_bowl_on_the_table/episode_1/cam_high/frame_000105.jpg", "images/droid_oxe/Move_the_bowl_on_the_table/episode_1/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Move_the_bowl_on_the_table/episode_1/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the bowl on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the bowl on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_black_cap_to_the_left_episode_1-train-bf_000210-af_000225-plus-0000227", "task": "Move the black cap to the left", "image": [ "images/droid_oxe/Move_the_black_cap_to_the_left/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_black_cap_to_the_left/episode_1/cam_high/frame_000337.jpg", "images/droid_oxe/Move_the_black_cap_to_the_left/episode_1/cam_high/frame_000210.jpg", "images/droid_oxe/Move_the_black_cap_to_the_left/episode_1/cam_left_wrist/frame_000210.jpg", "images/droid_oxe/Move_the_black_cap_to_the_left/episode_1/cam_right_wrist/frame_000210.jpg", "images/droid_oxe/Move_the_black_cap_to_the_left/episode_1/cam_high/frame_000225.jpg", "images/droid_oxe/Move_the_black_cap_to_the_left/episode_1/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Move_the_black_cap_to_the_left/episode_1/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the black cap to the left`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the black cap to the left`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_turquoise_object_in_the_pot_on_top_of_the_chest_drawer_episode_0-train-bf_000225-af_000285-plus-0000739", "task": "Put the turquoise object in the pot on top of the chest drawer", "image": [ "images/droid_oxe/Put_the_turquoise_object_in_the_pot_on_top_of_the_chest_drawer/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_turquoise_object_in_the_pot_on_top_of_the_chest_drawer/episode_0/cam_high/frame_000563.jpg", "images/droid_oxe/Put_the_turquoise_object_in_the_pot_on_top_of_the_chest_drawer/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Put_the_turquoise_object_in_the_pot_on_top_of_the_chest_drawer/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_turquoise_object_in_the_pot_on_top_of_the_chest_drawer/episode_0/cam_right_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_turquoise_object_in_the_pot_on_top_of_the_chest_drawer/episode_0/cam_high/frame_000285.jpg", "images/droid_oxe/Put_the_turquoise_object_in_the_pot_on_top_of_the_chest_drawer/episode_0/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/Put_the_turquoise_object_in_the_pot_on_top_of_the_chest_drawer/episode_0/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the turquoise object in the pot on top of the chest drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the turquoise object in the pot on top of the chest drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 18.2, "hop_absolute_value": 18.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Push_down_the_top_of_the_black_bottle_episode_0-train-bf_000030-af_000105-plus-0000096", "task": "Push down the top of the black bottle", "image": [ "images/droid_oxe/Push_down_the_top_of_the_black_bottle/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Push_down_the_top_of_the_black_bottle/episode_0/cam_high/frame_000167.jpg", "images/droid_oxe/Push_down_the_top_of_the_black_bottle/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Push_down_the_top_of_the_black_bottle/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Push_down_the_top_of_the_black_bottle/episode_0/cam_right_wrist/frame_000030.jpg", "images/droid_oxe/Push_down_the_top_of_the_black_bottle/episode_0/cam_high/frame_000105.jpg", "images/droid_oxe/Push_down_the_top_of_the_black_bottle/episode_0/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Push_down_the_top_of_the_black_bottle/episode_0/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push down the top of the black bottle`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push down the top of the black bottle`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 55.6, "hop_absolute_value": 55.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_book_in_the_black_box_on_the_table_episode_0-train-bf_000405-af_000420-plus-0000690", "task": "Put the book in the black box on the table", "image": [ "images/droid_oxe/Put_the_book_in_the_black_box_on_the_table/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_book_in_the_black_box_on_the_table/episode_0/cam_high/frame_000468.jpg", "images/droid_oxe/Put_the_book_in_the_black_box_on_the_table/episode_0/cam_high/frame_000405.jpg", "images/droid_oxe/Put_the_book_in_the_black_box_on_the_table/episode_0/cam_left_wrist/frame_000405.jpg", "images/droid_oxe/Put_the_book_in_the_black_box_on_the_table/episode_0/cam_right_wrist/frame_000405.jpg", "images/droid_oxe/Put_the_book_in_the_black_box_on_the_table/episode_0/cam_high/frame_000420.jpg", "images/droid_oxe/Put_the_book_in_the_black_box_on_the_table/episode_0/cam_left_wrist/frame_000420.jpg", "images/droid_oxe/Put_the_book_in_the_black_box_on_the_table/episode_0/cam_right_wrist/frame_000420.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the book in the black box on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the book in the black box on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pour_the_things_in_the_orange_cup_onto_the_plate_episode_1-train-bf_000015-af_000030-plus-0000272", "task": "Pour the things in the orange cup onto the plate", "image": [ "images/droid_oxe/Pour_the_things_in_the_orange_cup_onto_the_plate/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Pour_the_things_in_the_orange_cup_onto_the_plate/episode_1/cam_high/frame_000481.jpg", "images/droid_oxe/Pour_the_things_in_the_orange_cup_onto_the_plate/episode_1/cam_high/frame_000015.jpg", "images/droid_oxe/Pour_the_things_in_the_orange_cup_onto_the_plate/episode_1/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Pour_the_things_in_the_orange_cup_onto_the_plate/episode_1/cam_right_wrist/frame_000015.jpg", "images/droid_oxe/Pour_the_things_in_the_orange_cup_onto_the_plate/episode_1/cam_high/frame_000030.jpg", "images/droid_oxe/Pour_the_things_in_the_orange_cup_onto_the_plate/episode_1/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Pour_the_things_in_the_orange_cup_onto_the_plate/episode_1/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pour the things in the orange cup onto the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pour the things in the orange cup onto the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.2, "hop_absolute_value": 3.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_can_from_the_sink_and_put_it_on_the_counter_episode_3-train-bf_000180-af_000195-plus-0000144", "task": "Pick up the can from the sink and put it on the counter", "image": [ "images/droid_oxe/Pick_up_the_can_from_the_sink_and_put_it_on_the_counter/episode_5/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_can_from_the_sink_and_put_it_on_the_counter/episode_5/cam_high/frame_000196.jpg", "images/droid_oxe/Pick_up_the_can_from_the_sink_and_put_it_on_the_counter/episode_3/cam_high/frame_000180.jpg", "images/droid_oxe/Pick_up_the_can_from_the_sink_and_put_it_on_the_counter/episode_3/cam_left_wrist/frame_000180.jpg", "images/droid_oxe/Pick_up_the_can_from_the_sink_and_put_it_on_the_counter/episode_3/cam_right_wrist/frame_000180.jpg", "images/droid_oxe/Pick_up_the_can_from_the_sink_and_put_it_on_the_counter/episode_3/cam_high/frame_000195.jpg", "images/droid_oxe/Pick_up_the_can_from_the_sink_and_put_it_on_the_counter/episode_3/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Pick_up_the_can_from_the_sink_and_put_it_on_the_counter/episode_3/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the can from the sink and put it on the counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the can from the sink and put it on the counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Mix_the_contents_in_the_blue_plate_with_the_cake_knife_episode_0-train-bf_000075-af_000300-plus-0000404", "task": "Mix the contents in the blue plate with the cake knife", "image": [ "images/droid_oxe/Mix_the_contents_in_the_blue_plate_with_the_cake_knife/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Mix_the_contents_in_the_blue_plate_with_the_cake_knife/episode_0/cam_high/frame_000448.jpg", "images/droid_oxe/Mix_the_contents_in_the_blue_plate_with_the_cake_knife/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Mix_the_contents_in_the_blue_plate_with_the_cake_knife/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Mix_the_contents_in_the_blue_plate_with_the_cake_knife/episode_0/cam_right_wrist/frame_000075.jpg", "images/droid_oxe/Mix_the_contents_in_the_blue_plate_with_the_cake_knife/episode_0/cam_high/frame_000300.jpg", "images/droid_oxe/Mix_the_contents_in_the_blue_plate_with_the_cake_knife/episode_0/cam_left_wrist/frame_000300.jpg", "images/droid_oxe/Mix_the_contents_in_the_blue_plate_with_the_cake_knife/episode_0/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Mix the contents in the blue plate with the cake knife`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Mix the contents in the blue plate with the cake knife`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_object_to_the_right_episode_2-train-bf_000165-af_000240-plus-0000264", "task": "Move the object to the right", "image": [ "images/droid_oxe/Move_the_object_to_the_right/episode_6/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_object_to_the_right/episode_6/cam_high/frame_000201.jpg", "images/droid_oxe/Move_the_object_to_the_right/episode_2/cam_high/frame_000165.jpg", "images/droid_oxe/Move_the_object_to_the_right/episode_2/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Move_the_object_to_the_right/episode_2/cam_right_wrist/frame_000165.jpg", "images/droid_oxe/Move_the_object_to_the_right/episode_2/cam_high/frame_000240.jpg", "images/droid_oxe/Move_the_object_to_the_right/episode_2/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Move_the_object_to_the_right/episode_2/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the object to the right`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the object to the right`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_some_contents_from_the_bowl_and_put_them_inside_the_container_episode_0-train-bf_000120-af_000150-plus-0000266", "task": "Pick up some contents from the bowl and put them inside the container", "image": [ "images/droid_oxe/Pick_up_some_contents_from_the_bowl_and_put_them_inside_the_container/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_some_contents_from_the_bowl_and_put_them_inside_the_container/episode_0/cam_high/frame_000430.jpg", "images/droid_oxe/Pick_up_some_contents_from_the_bowl_and_put_them_inside_the_container/episode_0/cam_high/frame_000120.jpg", "images/droid_oxe/Pick_up_some_contents_from_the_bowl_and_put_them_inside_the_container/episode_0/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Pick_up_some_contents_from_the_bowl_and_put_them_inside_the_container/episode_0/cam_right_wrist/frame_000120.jpg", "images/droid_oxe/Pick_up_some_contents_from_the_bowl_and_put_them_inside_the_container/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Pick_up_some_contents_from_the_bowl_and_put_them_inside_the_container/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Pick_up_some_contents_from_the_bowl_and_put_them_inside_the_container/episode_0/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up some contents from the bowl and put them inside the container`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up some contents from the bowl and put them inside the container`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Disconnect_the_black_plug_from_the_socket_episode_0-train-bf_000015-af_000075-plus-0000064", "task": "Disconnect the black plug from the socket", "image": [ "images/droid_oxe/Disconnect_the_black_plug_from_the_socket/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Disconnect_the_black_plug_from_the_socket/episode_0/cam_high/frame_000138.jpg", "images/droid_oxe/Disconnect_the_black_plug_from_the_socket/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Disconnect_the_black_plug_from_the_socket/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Disconnect_the_black_plug_from_the_socket/episode_0/cam_right_wrist/frame_000015.jpg", "images/droid_oxe/Disconnect_the_black_plug_from_the_socket/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Disconnect_the_black_plug_from_the_socket/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Disconnect_the_black_plug_from_the_socket/episode_0/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disconnect the black plug from the socket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disconnect the black plug from the socket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Open_the_drawer_thats_nearest_to_the_wall_episode_0-train-bf_000060-af_000135-plus-0000107", "task": "Open the drawer thats nearest to the wall", "image": [ "images/droid_oxe/Open_the_drawer_thats_nearest_to_the_wall/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Open_the_drawer_thats_nearest_to_the_wall/episode_0/cam_high/frame_000174.jpg", "images/droid_oxe/Open_the_drawer_thats_nearest_to_the_wall/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Open_the_drawer_thats_nearest_to_the_wall/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Open_the_drawer_thats_nearest_to_the_wall/episode_0/cam_right_wrist/frame_000060.jpg", "images/droid_oxe/Open_the_drawer_thats_nearest_to_the_wall/episode_0/cam_high/frame_000135.jpg", "images/droid_oxe/Open_the_drawer_thats_nearest_to_the_wall/episode_0/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Open_the_drawer_thats_nearest_to_the_wall/episode_0/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the drawer thats nearest to the wall`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the drawer thats nearest to the wall`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Turn_the_stapler_then_pick_up_the_paper_from_the_table_and_staple_it_episode_0-train-bf_000000-af_000015-plus-0000289", "task": "Turn the stapler then pick up the paper from the table and staple it", "image": [ "images/droid_oxe/Turn_the_stapler_then_pick_up_the_paper_from_the_table_and_staple_it/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Turn_the_stapler_then_pick_up_the_paper_from_the_table_and_staple_it/episode_0/cam_high/frame_000562.jpg", "images/droid_oxe/Turn_the_stapler_then_pick_up_the_paper_from_the_table_and_staple_it/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Turn_the_stapler_then_pick_up_the_paper_from_the_table_and_staple_it/episode_0/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Turn_the_stapler_then_pick_up_the_paper_from_the_table_and_staple_it/episode_0/cam_right_wrist/frame_000000.jpg", "images/droid_oxe/Turn_the_stapler_then_pick_up_the_paper_from_the_table_and_staple_it/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Turn_the_stapler_then_pick_up_the_paper_from_the_table_and_staple_it/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Turn_the_stapler_then_pick_up_the_paper_from_the_table_and_staple_it/episode_0/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Turn the stapler then pick up the paper from the table and staple it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Turn the stapler then pick up the paper from the table and staple it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.7, "hop_absolute_value": 2.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_spoon_to_take_out_some_contents_from_the_red_bowl_and_place_them_in_the_black_pot_episode_0-train-bf_000045-af_000360-plus-0000448", "task": "Use the spoon to take out some contents from the red bowl and place them in the black pot", "image": [ "images/droid_oxe/Use_the_spoon_to_take_out_some_contents_from_the_red_bowl_and_place_them_in_the_black_pot/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_spoon_to_take_out_some_contents_from_the_red_bowl_and_place_them_in_the_black_pot/episode_0/cam_high/frame_000499.jpg", "images/droid_oxe/Use_the_spoon_to_take_out_some_contents_from_the_red_bowl_and_place_them_in_the_black_pot/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Use_the_spoon_to_take_out_some_contents_from_the_red_bowl_and_place_them_in_the_black_pot/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Use_the_spoon_to_take_out_some_contents_from_the_red_bowl_and_place_them_in_the_black_pot/episode_0/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Use_the_spoon_to_take_out_some_contents_from_the_red_bowl_and_place_them_in_the_black_pot/episode_0/cam_high/frame_000360.jpg", "images/droid_oxe/Use_the_spoon_to_take_out_some_contents_from_the_red_bowl_and_place_them_in_the_black_pot/episode_0/cam_left_wrist/frame_000360.jpg", "images/droid_oxe/Use_the_spoon_to_take_out_some_contents_from_the_red_bowl_and_place_them_in_the_black_pot/episode_0/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the spoon to take out some contents from the red bowl and place them in the black pot`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the spoon to take out some contents from the red bowl and place them in the black pot`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Open_the_right_drawer_of_the_cabinet_episode_1-train-bf_000120-af_000135-plus-0000082", "task": "Open the right drawer of the cabinet", "image": [ "images/droid_oxe/Open_the_right_drawer_of_the_cabinet/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Open_the_right_drawer_of_the_cabinet/episode_0/cam_high/frame_000210.jpg", "images/droid_oxe/Open_the_right_drawer_of_the_cabinet/episode_1/cam_high/frame_000120.jpg", "images/droid_oxe/Open_the_right_drawer_of_the_cabinet/episode_1/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Open_the_right_drawer_of_the_cabinet/episode_1/cam_right_wrist/frame_000120.jpg", "images/droid_oxe/Open_the_right_drawer_of_the_cabinet/episode_1/cam_high/frame_000135.jpg", "images/droid_oxe/Open_the_right_drawer_of_the_cabinet/episode_1/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Open_the_right_drawer_of_the_cabinet/episode_1/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the right drawer of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the right drawer of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Rotate_the_third_rod_backwards_episode_1-train-bf_000090-af_000105-plus-0000168", "task": "Rotate the third rod backwards", "image": [ "images/droid_oxe/Rotate_the_third_rod_backwards/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Rotate_the_third_rod_backwards/episode_0/cam_high/frame_000254.jpg", "images/droid_oxe/Rotate_the_third_rod_backwards/episode_1/cam_high/frame_000090.jpg", "images/droid_oxe/Rotate_the_third_rod_backwards/episode_1/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Rotate_the_third_rod_backwards/episode_1/cam_right_wrist/frame_000090.jpg", "images/droid_oxe/Rotate_the_third_rod_backwards/episode_1/cam_high/frame_000105.jpg", "images/droid_oxe/Rotate_the_third_rod_backwards/episode_1/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Rotate_the_third_rod_backwards/episode_1/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Rotate the third rod backwards`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Rotate the third rod backwards`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 8.3, "hop_absolute_value": 8.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_tea_towel_to_wipe_the_plate_episode_0-train-bf_000105-af_000585-plus-0000506", "task": "Use the tea towel to wipe the plate", "image": [ "images/droid_oxe/Use_the_tea_towel_to_wipe_the_plate/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_tea_towel_to_wipe_the_plate/episode_0/cam_high/frame_000714.jpg", "images/droid_oxe/Use_the_tea_towel_to_wipe_the_plate/episode_0/cam_high/frame_000105.jpg", "images/droid_oxe/Use_the_tea_towel_to_wipe_the_plate/episode_0/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Use_the_tea_towel_to_wipe_the_plate/episode_0/cam_right_wrist/frame_000105.jpg", "images/droid_oxe/Use_the_tea_towel_to_wipe_the_plate/episode_0/cam_high/frame_000585.jpg", "images/droid_oxe/Use_the_tea_towel_to_wipe_the_plate/episode_0/cam_left_wrist/frame_000585.jpg", "images/droid_oxe/Use_the_tea_towel_to_wipe_the_plate/episode_0/cam_right_wrist/frame_000585.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the tea towel to wipe the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the tea towel to wipe the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Turn_off_the_left_light_switch_episode_10-train-bf_000165-af_000465-plus-0000435", "task": "Turn off the left light switch", "image": [ "images/droid_oxe/Turn_off_the_left_light_switch/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Turn_off_the_left_light_switch/episode_1/cam_high/frame_000241.jpg", "images/droid_oxe/Turn_off_the_left_light_switch/episode_10/cam_high/frame_000165.jpg", "images/droid_oxe/Turn_off_the_left_light_switch/episode_10/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Turn_off_the_left_light_switch/episode_10/cam_right_wrist/frame_000165.jpg", "images/droid_oxe/Turn_off_the_left_light_switch/episode_10/cam_high/frame_000465.jpg", "images/droid_oxe/Turn_off_the_left_light_switch/episode_10/cam_left_wrist/frame_000465.jpg", "images/droid_oxe/Turn_off_the_left_light_switch/episode_10/cam_right_wrist/frame_000465.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Turn off the left light switch`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Turn off the left light switch`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Switch_off_the_extension_cable_episode_4-train-bf_000030-af_000060-plus-0000100", "task": "Switch off the extension cable", "image": [ "images/droid_oxe/Switch_off_the_extension_cable/episode_3/cam_high/frame_000000.jpg", "images/droid_oxe/Switch_off_the_extension_cable/episode_3/cam_high/frame_000176.jpg", "images/droid_oxe/Switch_off_the_extension_cable/episode_4/cam_high/frame_000030.jpg", "images/droid_oxe/Switch_off_the_extension_cable/episode_4/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Switch_off_the_extension_cable/episode_4/cam_right_wrist/frame_000030.jpg", "images/droid_oxe/Switch_off_the_extension_cable/episode_4/cam_high/frame_000060.jpg", "images/droid_oxe/Switch_off_the_extension_cable/episode_4/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Switch_off_the_extension_cable/episode_4/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Switch off the extension cable`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Switch off the extension cable`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 18.2, "hop_absolute_value": 18.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_objects_in_the_plastic_and_put_them_on_the_table_episode_3-train-bf_000105-af_000120-plus-0000650", "task": "Remove the objects in the plastic and put them on the table", "image": [ "images/droid_oxe/Remove_the_objects_in_the_plastic_and_put_them_on_the_table/episode_3/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_objects_in_the_plastic_and_put_them_on_the_table/episode_3/cam_high/frame_000824.jpg", "images/droid_oxe/Remove_the_objects_in_the_plastic_and_put_them_on_the_table/episode_3/cam_high/frame_000105.jpg", "images/droid_oxe/Remove_the_objects_in_the_plastic_and_put_them_on_the_table/episode_3/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Remove_the_objects_in_the_plastic_and_put_them_on_the_table/episode_3/cam_right_wrist/frame_000105.jpg", "images/droid_oxe/Remove_the_objects_in_the_plastic_and_put_them_on_the_table/episode_3/cam_high/frame_000120.jpg", "images/droid_oxe/Remove_the_objects_in_the_plastic_and_put_them_on_the_table/episode_3/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Remove_the_objects_in_the_plastic_and_put_them_on_the_table/episode_3/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the objects in the plastic and put them on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the objects in the plastic and put them on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.1, "hop_absolute_value": 2.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_maroon_mug_to_the_left_episode_1-train-bf_000045-af_000090-plus-0000044", "task": "Move the maroon mug to the left", "image": [ "images/droid_oxe/Move_the_maroon_mug_to_the_left/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_maroon_mug_to_the_left/episode_1/cam_high/frame_000109.jpg", "images/droid_oxe/Move_the_maroon_mug_to_the_left/episode_1/cam_high/frame_000045.jpg", "images/droid_oxe/Move_the_maroon_mug_to_the_left/episode_1/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Move_the_maroon_mug_to_the_left/episode_1/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Move_the_maroon_mug_to_the_left/episode_1/cam_high/frame_000090.jpg", "images/droid_oxe/Move_the_maroon_mug_to_the_left/episode_1/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Move_the_maroon_mug_to_the_left/episode_1/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the maroon mug to the left`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the maroon mug to the left`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Get_the_red_object_from_the_black_pot_and_place_it_on_the_grey_plate_episode_0-train-bf_000045-af_000060-plus-0000410", "task": "Get the red object from the black pot and place it on the grey plate", "image": [ "images/droid_oxe/Get_the_red_object_from_the_black_pot_and_place_it_on_the_grey_plate/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Get_the_red_object_from_the_black_pot_and_place_it_on_the_grey_plate/episode_0/cam_high/frame_000317.jpg", "images/droid_oxe/Get_the_red_object_from_the_black_pot_and_place_it_on_the_grey_plate/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Get_the_red_object_from_the_black_pot_and_place_it_on_the_grey_plate/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Get_the_red_object_from_the_black_pot_and_place_it_on_the_grey_plate/episode_0/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Get_the_red_object_from_the_black_pot_and_place_it_on_the_grey_plate/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Get_the_red_object_from_the_black_pot_and_place_it_on_the_grey_plate/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Get_the_red_object_from_the_black_pot_and_place_it_on_the_grey_plate/episode_0/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Get the red object from the black pot and place it on the grey plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Get the red object from the black pot and place it on the grey plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.6, "hop_absolute_value": 5.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_lid_on_the_table_then_put_it_back_on_the_silver_container_episode_0-train-bf_000120-af_000375-plus-0000401", "task": "Put the lid on the table then put it back on the silver container", "image": [ "images/droid_oxe/Put_the_lid_on_the_table_then_put_it_back_on_the_silver_container/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_lid_on_the_table_then_put_it_back_on_the_silver_container/episode_0/cam_high/frame_000606.jpg", "images/droid_oxe/Put_the_lid_on_the_table_then_put_it_back_on_the_silver_container/episode_0/cam_high/frame_000120.jpg", "images/droid_oxe/Put_the_lid_on_the_table_then_put_it_back_on_the_silver_container/episode_0/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Put_the_lid_on_the_table_then_put_it_back_on_the_silver_container/episode_0/cam_right_wrist/frame_000120.jpg", "images/droid_oxe/Put_the_lid_on_the_table_then_put_it_back_on_the_silver_container/episode_0/cam_high/frame_000375.jpg", "images/droid_oxe/Put_the_lid_on_the_table_then_put_it_back_on_the_silver_container/episode_0/cam_left_wrist/frame_000375.jpg", "images/droid_oxe/Put_the_lid_on_the_table_then_put_it_back_on_the_silver_container/episode_0/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the lid on the table then put it back on the silver container`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the lid on the table then put it back on the silver container`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 53.1, "hop_absolute_value": 53.1, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_fork_from_the_blue_rack_episode_0-train-bf_000030-af_000150-plus-0000229", "task": "Remove the fork from the blue rack", "image": [ "images/droid_oxe/Remove_the_fork_from_the_blue_rack/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_fork_from_the_blue_rack/episode_0/cam_high/frame_000279.jpg", "images/droid_oxe/Remove_the_fork_from_the_blue_rack/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Remove_the_fork_from_the_blue_rack/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Remove_the_fork_from_the_blue_rack/episode_0/cam_right_wrist/frame_000030.jpg", "images/droid_oxe/Remove_the_fork_from_the_blue_rack/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Remove_the_fork_from_the_blue_rack/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Remove_the_fork_from_the_blue_rack/episode_0/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the fork from the blue rack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the fork from the blue rack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Close_the_microwave_door_episode_9-train-bf_000075-af_000135-plus-0000116", "task": "Close the microwave door", "image": [ "images/droid_oxe/Close_the_microwave_door/episode_6/cam_high/frame_000000.jpg", "images/droid_oxe/Close_the_microwave_door/episode_6/cam_high/frame_000157.jpg", "images/droid_oxe/Close_the_microwave_door/episode_9/cam_high/frame_000075.jpg", "images/droid_oxe/Close_the_microwave_door/episode_9/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Close_the_microwave_door/episode_9/cam_right_wrist/frame_000075.jpg", "images/droid_oxe/Close_the_microwave_door/episode_9/cam_high/frame_000135.jpg", "images/droid_oxe/Close_the_microwave_door/episode_9/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Close_the_microwave_door/episode_9/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the microwave door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the microwave door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.1, "hop_absolute_value": 57.1, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Push_the_left_lever_down_episode_1-train-bf_000045-af_000105-plus-0000060", "task": "Push the left lever down", "image": [ "images/droid_oxe/Push_the_left_lever_down/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Push_the_left_lever_down/episode_1/cam_high/frame_000133.jpg", "images/droid_oxe/Push_the_left_lever_down/episode_1/cam_high/frame_000045.jpg", "images/droid_oxe/Push_the_left_lever_down/episode_1/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Push_the_left_lever_down/episode_1/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Push_the_left_lever_down/episode_1/cam_high/frame_000105.jpg", "images/droid_oxe/Push_the_left_lever_down/episode_1/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Push_the_left_lever_down/episode_1/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the left lever down`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the left lever down`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_light_blue_cup_to_the_right_move_the_green_bowl_to_the_right_stir_the_contents_in_the_bowl_using_the_white_spoo_episode_0-train-bf_000225-af_000360-plus-0000365", "task": "Move the light blue cup to the right move the green bowl to the right stir the contents in the bowl using the white spoo", "image": [ "images/droid_oxe/Move_the_light_blue_cup_to_the_right_move_the_green_bowl_to_the_right_stir_the_contents_in_the_bowl_using_the_white_spoo/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_light_blue_cup_to_the_right_move_the_green_bowl_to_the_right_stir_the_contents_in_the_bowl_using_the_white_spoo/episode_0/cam_high/frame_000524.jpg", "images/droid_oxe/Move_the_light_blue_cup_to_the_right_move_the_green_bowl_to_the_right_stir_the_contents_in_the_bowl_using_the_white_spoo/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Move_the_light_blue_cup_to_the_right_move_the_green_bowl_to_the_right_stir_the_contents_in_the_bowl_using_the_white_spoo/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Move_the_light_blue_cup_to_the_right_move_the_green_bowl_to_the_right_stir_the_contents_in_the_bowl_using_the_white_spoo/episode_0/cam_right_wrist/frame_000225.jpg", "images/droid_oxe/Move_the_light_blue_cup_to_the_right_move_the_green_bowl_to_the_right_stir_the_contents_in_the_bowl_using_the_white_spoo/episode_0/cam_high/frame_000360.jpg", "images/droid_oxe/Move_the_light_blue_cup_to_the_right_move_the_green_bowl_to_the_right_stir_the_contents_in_the_bowl_using_the_white_spoo/episode_0/cam_left_wrist/frame_000360.jpg", "images/droid_oxe/Move_the_light_blue_cup_to_the_right_move_the_green_bowl_to_the_right_stir_the_contents_in_the_bowl_using_the_white_spoo/episode_0/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the light blue cup to the right move the green bowl to the right stir the contents in the bowl using the white spoo`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the light blue cup to the right move the green bowl to the right stir the contents in the bowl using the white spoo`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 47.4, "hop_absolute_value": 47.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Tip_the_mug_over_the_coffee_machine_chamber_episode_0-train-bf_000150-af_000165-plus-0000137", "task": "Tip the mug over the coffee machine chamber", "image": [ "images/droid_oxe/Tip_the_mug_over_the_coffee_machine_chamber/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Tip_the_mug_over_the_coffee_machine_chamber/episode_0/cam_high/frame_000237.jpg", "images/droid_oxe/Tip_the_mug_over_the_coffee_machine_chamber/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Tip_the_mug_over_the_coffee_machine_chamber/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Tip_the_mug_over_the_coffee_machine_chamber/episode_0/cam_right_wrist/frame_000150.jpg", "images/droid_oxe/Tip_the_mug_over_the_coffee_machine_chamber/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Tip_the_mug_over_the_coffee_machine_chamber/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Tip_the_mug_over_the_coffee_machine_chamber/episode_0/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Tip the mug over the coffee machine chamber`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Tip the mug over the coffee machine chamber`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_wooden_spatula_to_mix_the_contents_in_the_green_bowl_episode_0-train-bf_000285-af_000300-plus-0000550", "task": "Use the wooden spatula to mix the contents in the green bowl", "image": [ "images/droid_oxe/Use_the_wooden_spatula_to_mix_the_contents_in_the_green_bowl/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_wooden_spatula_to_mix_the_contents_in_the_green_bowl/episode_0/cam_high/frame_000432.jpg", "images/droid_oxe/Use_the_wooden_spatula_to_mix_the_contents_in_the_green_bowl/episode_0/cam_high/frame_000285.jpg", "images/droid_oxe/Use_the_wooden_spatula_to_mix_the_contents_in_the_green_bowl/episode_0/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/Use_the_wooden_spatula_to_mix_the_contents_in_the_green_bowl/episode_0/cam_right_wrist/frame_000285.jpg", "images/droid_oxe/Use_the_wooden_spatula_to_mix_the_contents_in_the_green_bowl/episode_0/cam_high/frame_000300.jpg", "images/droid_oxe/Use_the_wooden_spatula_to_mix_the_contents_in_the_green_bowl/episode_0/cam_left_wrist/frame_000300.jpg", "images/droid_oxe/Use_the_wooden_spatula_to_mix_the_contents_in_the_green_bowl/episode_0/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the wooden spatula to mix the contents in the green bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the wooden spatula to mix the contents in the green bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_the_cup_and_put_it_on_the_counter_episode_0-train-bf_000060-af_000195-plus-0000208", "task": "Pick the cup and put it on the counter", "image": [ "images/droid_oxe/Pick_the_cup_and_put_it_on_the_counter/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_the_cup_and_put_it_on_the_counter/episode_1/cam_high/frame_000220.jpg", "images/droid_oxe/Pick_the_cup_and_put_it_on_the_counter/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Pick_the_cup_and_put_it_on_the_counter/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Pick_the_cup_and_put_it_on_the_counter/episode_0/cam_right_wrist/frame_000060.jpg", "images/droid_oxe/Pick_the_cup_and_put_it_on_the_counter/episode_0/cam_high/frame_000195.jpg", "images/droid_oxe/Pick_the_cup_and_put_it_on_the_counter/episode_0/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Pick_the_cup_and_put_it_on_the_counter/episode_0/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the cup and put it on the counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the cup and put it on the counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 81.8, "hop_absolute_value": 81.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Push_the_lever_on_the_bread_toaster_down_episode_1-train-bf_000030-af_000135-plus-0000226", "task": "Push the lever on the bread toaster down", "image": [ "images/droid_oxe/Push_the_lever_on_the_bread_toaster_down/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Push_the_lever_on_the_bread_toaster_down/episode_0/cam_high/frame_000220.jpg", "images/droid_oxe/Push_the_lever_on_the_bread_toaster_down/episode_1/cam_high/frame_000030.jpg", "images/droid_oxe/Push_the_lever_on_the_bread_toaster_down/episode_1/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Push_the_lever_on_the_bread_toaster_down/episode_1/cam_right_wrist/frame_000030.jpg", "images/droid_oxe/Push_the_lever_on_the_bread_toaster_down/episode_1/cam_high/frame_000135.jpg", "images/droid_oxe/Push_the_lever_on_the_bread_toaster_down/episode_1/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Push_the_lever_on_the_bread_toaster_down/episode_1/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the lever on the bread toaster down`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the lever on the bread toaster down`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 41.2, "hop_absolute_value": 41.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_shirts_on_the_chair_onto_the_box_episode_1-train-bf_000330-af_000000-minus-0000030", "task": "Put the shirts on the chair onto the box", "image": [ "images/droid_oxe/Put_the_shirts_on_the_chair_onto_the_box/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_shirts_on_the_chair_onto_the_box/episode_1/cam_high/frame_000398.jpg", "images/droid_oxe/Put_the_shirts_on_the_chair_onto_the_box/episode_1/cam_high/frame_000330.jpg", "images/droid_oxe/Put_the_shirts_on_the_chair_onto_the_box/episode_1/cam_left_wrist/frame_000330.jpg", "images/droid_oxe/Put_the_shirts_on_the_chair_onto_the_box/episode_1/cam_right_wrist/frame_000330.jpg", "images/droid_oxe/Put_the_shirts_on_the_chair_onto_the_box/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_shirts_on_the_chair_onto_the_box/episode_1/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Put_the_shirts_on_the_chair_onto_the_box/episode_1/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the shirts on the chair onto the box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the shirts on the chair onto the box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_bottle_on_the_box_episode_6-train-bf_000060-af_000045-minus-0000031", "task": "Put the bottle on the box", "image": [ "images/droid_oxe/Put_the_bottle_on_the_box/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_bottle_on_the_box/episode_1/cam_high/frame_000180.jpg", "images/droid_oxe/Put_the_bottle_on_the_box/episode_6/cam_high/frame_000060.jpg", "images/droid_oxe/Put_the_bottle_on_the_box/episode_6/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Put_the_bottle_on_the_box/episode_6/cam_right_wrist/frame_000060.jpg", "images/droid_oxe/Put_the_bottle_on_the_box/episode_6/cam_high/frame_000045.jpg", "images/droid_oxe/Put_the_bottle_on_the_box/episode_6/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Put_the_bottle_on_the_box/episode_6/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the bottle on the box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the bottle on the box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_lid_from_the_pot_and_put_it_inside_the_bowl_episode_0-train-bf_000165-af_000120-minus-0000128", "task": "Remove the lid from the pot and put it inside the bowl", "image": [ "images/droid_oxe/Remove_the_lid_from_the_pot_and_put_it_inside_the_bowl/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_lid_from_the_pot_and_put_it_inside_the_bowl/episode_0/cam_high/frame_000263.jpg", "images/droid_oxe/Remove_the_lid_from_the_pot_and_put_it_inside_the_bowl/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Remove_the_lid_from_the_pot_and_put_it_inside_the_bowl/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Remove_the_lid_from_the_pot_and_put_it_inside_the_bowl/episode_0/cam_right_wrist/frame_000165.jpg", "images/droid_oxe/Remove_the_lid_from_the_pot_and_put_it_inside_the_bowl/episode_0/cam_high/frame_000120.jpg", "images/droid_oxe/Remove_the_lid_from_the_pot_and_put_it_inside_the_bowl/episode_0/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Remove_the_lid_from_the_pot_and_put_it_inside_the_bowl/episode_0/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the lid from the pot and put it inside the bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the lid from the pot and put it inside the bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -27.3, "hop_absolute_value": 27.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Take_the_yellow_block_and_put_it_in_the_box_episode_0-train-bf_000105-af_000015-minus-0000012", "task": "Take the yellow block and put it in the box", "image": [ "images/droid_oxe/Take_the_yellow_block_and_put_it_in_the_box/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Take_the_yellow_block_and_put_it_in_the_box/episode_0/cam_high/frame_000191.jpg", "images/droid_oxe/Take_the_yellow_block_and_put_it_in_the_box/episode_0/cam_high/frame_000105.jpg", "images/droid_oxe/Take_the_yellow_block_and_put_it_in_the_box/episode_0/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Take_the_yellow_block_and_put_it_in_the_box/episode_0/cam_right_wrist/frame_000105.jpg", "images/droid_oxe/Take_the_yellow_block_and_put_it_in_the_box/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Take_the_yellow_block_and_put_it_in_the_box/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Take_the_yellow_block_and_put_it_in_the_box/episode_0/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the yellow block and put it in the box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the yellow block and put it in the box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_green_towel_to_wipe_the_counter_episode_0-train-bf_000225-af_000060-minus-0000050", "task": "Use the green towel to wipe the counter", "image": [ "images/droid_oxe/Use_the_green_towel_to_wipe_the_counter/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_green_towel_to_wipe_the_counter/episode_0/cam_high/frame_000247.jpg", "images/droid_oxe/Use_the_green_towel_to_wipe_the_counter/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Use_the_green_towel_to_wipe_the_counter/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Use_the_green_towel_to_wipe_the_counter/episode_0/cam_right_wrist/frame_000225.jpg", "images/droid_oxe/Use_the_green_towel_to_wipe_the_counter/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Use_the_green_towel_to_wipe_the_counter/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Use_the_green_towel_to_wipe_the_counter/episode_0/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the green towel to wipe the counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the green towel to wipe the counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.3, "hop_absolute_value": 73.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_nail_polish_bottles_in_the_bag_episode_2-train-bf_000450-af_000240-minus-0000208", "task": "Put the nail polish bottles in the bag", "image": [ "images/droid_oxe/Put_the_nail_polish_bottles_in_the_bag/episode_2/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_nail_polish_bottles_in_the_bag/episode_2/cam_high/frame_001069.jpg", "images/droid_oxe/Put_the_nail_polish_bottles_in_the_bag/episode_2/cam_high/frame_000450.jpg", "images/droid_oxe/Put_the_nail_polish_bottles_in_the_bag/episode_2/cam_left_wrist/frame_000450.jpg", "images/droid_oxe/Put_the_nail_polish_bottles_in_the_bag/episode_2/cam_right_wrist/frame_000450.jpg", "images/droid_oxe/Put_the_nail_polish_bottles_in_the_bag/episode_2/cam_high/frame_000240.jpg", "images/droid_oxe/Put_the_nail_polish_bottles_in_the_bag/episode_2/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Put_the_nail_polish_bottles_in_the_bag/episode_2/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the nail polish bottles in the bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the nail polish bottles in the bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -46.7, "hop_absolute_value": 46.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_rectangular_block_on_the_object_to_the_right_episode_0-train-bf_000139-af_000120-minus-0000044", "task": "Move the rectangular block on the object to the right", "image": [ "images/droid_oxe/Move_the_rectangular_block_on_the_object_to_the_right/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_rectangular_block_on_the_object_to_the_right/episode_0/cam_high/frame_000139.jpg", "images/droid_oxe/Move_the_rectangular_block_on_the_object_to_the_right/episode_0/cam_high/frame_000139.jpg", "images/droid_oxe/Move_the_rectangular_block_on_the_object_to_the_right/episode_0/cam_left_wrist/frame_000139.jpg", "images/droid_oxe/Move_the_rectangular_block_on_the_object_to_the_right/episode_0/cam_right_wrist/frame_000139.jpg", "images/droid_oxe/Move_the_rectangular_block_on_the_object_to_the_right/episode_0/cam_high/frame_000120.jpg", "images/droid_oxe/Move_the_rectangular_block_on_the_object_to_the_right/episode_0/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Move_the_rectangular_block_on_the_object_to_the_right/episode_0/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the rectangular block on the object to the right`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the rectangular block on the object to the right`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_all_the_objects_surrounding_the_white_plate_and_arrange_them_on_the_plate_episode_3-train-bf_000600-af_000360-minus-0000228", "task": "Pick up all the objects surrounding the white plate and arrange them on the plate", "image": [ "images/droid_oxe/Pick_up_all_the_objects_surrounding_the_white_plate_and_arrange_them_on_the_plate/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_all_the_objects_surrounding_the_white_plate_and_arrange_them_on_the_plate/episode_1/cam_high/frame_000984.jpg", "images/droid_oxe/Pick_up_all_the_objects_surrounding_the_white_plate_and_arrange_them_on_the_plate/episode_3/cam_high/frame_000600.jpg", "images/droid_oxe/Pick_up_all_the_objects_surrounding_the_white_plate_and_arrange_them_on_the_plate/episode_3/cam_left_wrist/frame_000600.jpg", "images/droid_oxe/Pick_up_all_the_objects_surrounding_the_white_plate_and_arrange_them_on_the_plate/episode_3/cam_right_wrist/frame_000600.jpg", "images/droid_oxe/Pick_up_all_the_objects_surrounding_the_white_plate_and_arrange_them_on_the_plate/episode_3/cam_high/frame_000360.jpg", "images/droid_oxe/Pick_up_all_the_objects_surrounding_the_white_plate_and_arrange_them_on_the_plate/episode_3/cam_left_wrist/frame_000360.jpg", "images/droid_oxe/Pick_up_all_the_objects_surrounding_the_white_plate_and_arrange_them_on_the_plate/episode_3/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up all the objects surrounding the white plate and arrange them on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up all the objects surrounding the white plate and arrange them on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_some_snacks_on_the_white_plate_from_the_pack_episode_0-train-bf_000255-af_000180-minus-0000147", "task": "Put some snacks on the white plate from the pack", "image": [ "images/droid_oxe/Put_some_snacks_on_the_white_plate_from_the_pack/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_some_snacks_on_the_white_plate_from_the_pack/episode_0/cam_high/frame_000305.jpg", "images/droid_oxe/Put_some_snacks_on_the_white_plate_from_the_pack/episode_0/cam_high/frame_000255.jpg", "images/droid_oxe/Put_some_snacks_on_the_white_plate_from_the_pack/episode_0/cam_left_wrist/frame_000255.jpg", "images/droid_oxe/Put_some_snacks_on_the_white_plate_from_the_pack/episode_0/cam_right_wrist/frame_000255.jpg", "images/droid_oxe/Put_some_snacks_on_the_white_plate_from_the_pack/episode_0/cam_high/frame_000180.jpg", "images/droid_oxe/Put_some_snacks_on_the_white_plate_from_the_pack/episode_0/cam_left_wrist/frame_000180.jpg", "images/droid_oxe/Put_some_snacks_on_the_white_plate_from_the_pack/episode_0/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put some snacks on the white plate from the pack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put some snacks on the white plate from the pack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -29.4, "hop_absolute_value": 29.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Take_the_lid_and_put_it_on_the_bottle_episode_1-train-bf_000120-af_000075-minus-0000036", "task": "Take the lid and put it on the bottle", "image": [ "images/droid_oxe/Take_the_lid_and_put_it_on_the_bottle/episode_2/cam_high/frame_000000.jpg", "images/droid_oxe/Take_the_lid_and_put_it_on_the_bottle/episode_2/cam_high/frame_000274.jpg", "images/droid_oxe/Take_the_lid_and_put_it_on_the_bottle/episode_1/cam_high/frame_000120.jpg", "images/droid_oxe/Take_the_lid_and_put_it_on_the_bottle/episode_1/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Take_the_lid_and_put_it_on_the_bottle/episode_1/cam_right_wrist/frame_000120.jpg", "images/droid_oxe/Take_the_lid_and_put_it_on_the_bottle/episode_1/cam_high/frame_000075.jpg", "images/droid_oxe/Take_the_lid_and_put_it_on_the_bottle/episode_1/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Take_the_lid_and_put_it_on_the_bottle/episode_1/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the lid and put it on the bottle`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the lid and put it on the bottle`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -37.5, "hop_absolute_value": 37.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Plug_the_charger_head_into_the_adapter_episode_2-train-bf_000450-af_000285-minus-0000222", "task": "Plug the charger head into the adapter", "image": [ "images/droid_oxe/Plug_the_charger_head_into_the_adapter/episode_2/cam_high/frame_000000.jpg", "images/droid_oxe/Plug_the_charger_head_into_the_adapter/episode_2/cam_high/frame_000525.jpg", "images/droid_oxe/Plug_the_charger_head_into_the_adapter/episode_2/cam_high/frame_000450.jpg", "images/droid_oxe/Plug_the_charger_head_into_the_adapter/episode_2/cam_left_wrist/frame_000450.jpg", "images/droid_oxe/Plug_the_charger_head_into_the_adapter/episode_2/cam_right_wrist/frame_000450.jpg", "images/droid_oxe/Plug_the_charger_head_into_the_adapter/episode_2/cam_high/frame_000285.jpg", "images/droid_oxe/Plug_the_charger_head_into_the_adapter/episode_2/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/Plug_the_charger_head_into_the_adapter/episode_2/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Plug the charger head into the adapter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Plug the charger head into the adapter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -36.7, "hop_absolute_value": 36.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_spoon_and_put_it_in_the_bowl_episode_1-train-bf_000045-af_000000-minus-0000001", "task": "Pick up the spoon and put it in the bowl", "image": [ "images/droid_oxe/Pick_up_the_spoon_and_put_it_in_the_bowl/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_spoon_and_put_it_in_the_bowl/episode_1/cam_high/frame_000186.jpg", "images/droid_oxe/Pick_up_the_spoon_and_put_it_in_the_bowl/episode_1/cam_high/frame_000045.jpg", "images/droid_oxe/Pick_up_the_spoon_and_put_it_in_the_bowl/episode_1/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Pick_up_the_spoon_and_put_it_in_the_bowl/episode_1/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Pick_up_the_spoon_and_put_it_in_the_bowl/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_spoon_and_put_it_in_the_bowl/episode_1/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Pick_up_the_spoon_and_put_it_in_the_bowl/episode_1/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the spoon and put it in the bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the spoon and put it in the bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_blocks_in_the_bowl_episode_4-train-bf_000090-af_000060-minus-0000043", "task": "Put the blocks in the bowl", "image": [ "images/droid_oxe/Put_the_blocks_in_the_bowl/episode_12/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_blocks_in_the_bowl/episode_12/cam_high/frame_000955.jpg", "images/droid_oxe/Put_the_blocks_in_the_bowl/episode_4/cam_high/frame_000090.jpg", "images/droid_oxe/Put_the_blocks_in_the_bowl/episode_4/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Put_the_blocks_in_the_bowl/episode_4/cam_right_wrist/frame_000090.jpg", "images/droid_oxe/Put_the_blocks_in_the_bowl/episode_4/cam_high/frame_000060.jpg", "images/droid_oxe/Put_the_blocks_in_the_bowl/episode_4/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Put_the_blocks_in_the_bowl/episode_4/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the blocks in the bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the blocks in the bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Take_the_mug_from_the_sink_and_put_it_on_the_dish_rack_episode_0-train-bf_000480-af_000105-minus-0000098", "task": "Take the mug from the sink and put it on the dish rack", "image": [ "images/droid_oxe/Take_the_mug_from_the_sink_and_put_it_on_the_dish_rack/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Take_the_mug_from_the_sink_and_put_it_on_the_dish_rack/episode_0/cam_high/frame_000503.jpg", "images/droid_oxe/Take_the_mug_from_the_sink_and_put_it_on_the_dish_rack/episode_0/cam_high/frame_000480.jpg", "images/droid_oxe/Take_the_mug_from_the_sink_and_put_it_on_the_dish_rack/episode_0/cam_left_wrist/frame_000480.jpg", "images/droid_oxe/Take_the_mug_from_the_sink_and_put_it_on_the_dish_rack/episode_0/cam_right_wrist/frame_000480.jpg", "images/droid_oxe/Take_the_mug_from_the_sink_and_put_it_on_the_dish_rack/episode_0/cam_high/frame_000105.jpg", "images/droid_oxe/Take_the_mug_from_the_sink_and_put_it_on_the_dish_rack/episode_0/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Take_the_mug_from_the_sink_and_put_it_on_the_dish_rack/episode_0/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the mug from the sink and put it on the dish rack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the mug from the sink and put it on the dish rack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -78.1, "hop_absolute_value": 78.1, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_tissue_to_wipe_the_book_episode_1-train-bf_000105-af_000000-minus-0000005", "task": "Use the tissue to wipe the book", "image": [ "images/droid_oxe/Use_the_tissue_to_wipe_the_book/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_tissue_to_wipe_the_book/episode_0/cam_high/frame_000372.jpg", "images/droid_oxe/Use_the_tissue_to_wipe_the_book/episode_1/cam_high/frame_000105.jpg", "images/droid_oxe/Use_the_tissue_to_wipe_the_book/episode_1/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Use_the_tissue_to_wipe_the_book/episode_1/cam_right_wrist/frame_000105.jpg", "images/droid_oxe/Use_the_tissue_to_wipe_the_book/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_tissue_to_wipe_the_book/episode_1/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Use_the_tissue_to_wipe_the_book/episode_1/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the tissue to wipe the book`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the tissue to wipe the book`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_grey_remote_and_put_it_on_the_table_episode_0-train-bf_000375-af_000015-minus-0000026", "task": "Pick up the grey remote and put it on the table", "image": [ "images/droid_oxe/Pick_up_the_grey_remote_and_put_it_on_the_table/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_grey_remote_and_put_it_on_the_table/episode_0/cam_high/frame_000617.jpg", "images/droid_oxe/Pick_up_the_grey_remote_and_put_it_on_the_table/episode_0/cam_high/frame_000375.jpg", "images/droid_oxe/Pick_up_the_grey_remote_and_put_it_on_the_table/episode_0/cam_left_wrist/frame_000375.jpg", "images/droid_oxe/Pick_up_the_grey_remote_and_put_it_on_the_table/episode_0/cam_right_wrist/frame_000375.jpg", "images/droid_oxe/Pick_up_the_grey_remote_and_put_it_on_the_table/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Pick_up_the_grey_remote_and_put_it_on_the_table/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Pick_up_the_grey_remote_and_put_it_on_the_table/episode_0/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the grey remote and put it on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the grey remote and put it on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -96.0, "hop_absolute_value": 96.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pour_the_contents_in_the_mug_cup_inside_the_clear_cup_episode_0-train-bf_000690-af_000570-minus-0000843", "task": "Pour the contents in the mug cup inside the clear cup", "image": [ "images/droid_oxe/Pour_the_contents_in_the_mug_cup_inside_the_clear_cup/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pour_the_contents_in_the_mug_cup_inside_the_clear_cup/episode_0/cam_high/frame_000773.jpg", "images/droid_oxe/Pour_the_contents_in_the_mug_cup_inside_the_clear_cup/episode_0/cam_high/frame_000690.jpg", "images/droid_oxe/Pour_the_contents_in_the_mug_cup_inside_the_clear_cup/episode_0/cam_left_wrist/frame_000690.jpg", "images/droid_oxe/Pour_the_contents_in_the_mug_cup_inside_the_clear_cup/episode_0/cam_right_wrist/frame_000690.jpg", "images/droid_oxe/Pour_the_contents_in_the_mug_cup_inside_the_clear_cup/episode_0/cam_high/frame_000570.jpg", "images/droid_oxe/Pour_the_contents_in_the_mug_cup_inside_the_clear_cup/episode_0/cam_left_wrist/frame_000570.jpg", "images/droid_oxe/Pour_the_contents_in_the_mug_cup_inside_the_clear_cup/episode_0/cam_right_wrist/frame_000570.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pour the contents in the mug cup inside the clear cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pour the contents in the mug cup inside the clear cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -17.4, "hop_absolute_value": 17.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_sponge_from_the_pot_and_put_it_on_the_table_then_pick_up_the_lid_and_put_it_on_the_pot_episode_0-train-bf_000255-af_000000-minus-0000018", "task": "Pick up the sponge from the pot and put it on the table then pick up the lid and put it on the pot", "image": [ "images/droid_oxe/Pick_up_the_sponge_from_the_pot_and_put_it_on_the_table_then_pick_up_the_lid_and_put_it_on_the_pot/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_sponge_from_the_pot_and_put_it_on_the_table_then_pick_up_the_lid_and_put_it_on_the_pot/episode_0/cam_high/frame_000281.jpg", "images/droid_oxe/Pick_up_the_sponge_from_the_pot_and_put_it_on_the_table_then_pick_up_the_lid_and_put_it_on_the_pot/episode_0/cam_high/frame_000255.jpg", "images/droid_oxe/Pick_up_the_sponge_from_the_pot_and_put_it_on_the_table_then_pick_up_the_lid_and_put_it_on_the_pot/episode_0/cam_left_wrist/frame_000255.jpg", "images/droid_oxe/Pick_up_the_sponge_from_the_pot_and_put_it_on_the_table_then_pick_up_the_lid_and_put_it_on_the_pot/episode_0/cam_right_wrist/frame_000255.jpg", "images/droid_oxe/Pick_up_the_sponge_from_the_pot_and_put_it_on_the_table_then_pick_up_the_lid_and_put_it_on_the_pot/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_sponge_from_the_pot_and_put_it_on_the_table_then_pick_up_the_lid_and_put_it_on_the_pot/episode_0/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Pick_up_the_sponge_from_the_pot_and_put_it_on_the_table_then_pick_up_the_lid_and_put_it_on_the_pot/episode_0/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the sponge from the pot and put it on the table then pick up the lid and put it on the pot`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the sponge from the pot and put it on the table then pick up the lid and put it on the pot`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_gray_towel_from_the_carrier_bag_and_put_it_on_the_handle_of_the_couch_episode_0-train-bf_000214-af_000075-minus-0000047", "task": "Remove the gray towel from the carrier bag and put it on the handle of the couch", "image": [ "images/droid_oxe/Remove_the_gray_towel_from_the_carrier_bag_and_put_it_on_the_handle_of_the_couch/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_gray_towel_from_the_carrier_bag_and_put_it_on_the_handle_of_the_couch/episode_0/cam_high/frame_000214.jpg", "images/droid_oxe/Remove_the_gray_towel_from_the_carrier_bag_and_put_it_on_the_handle_of_the_couch/episode_0/cam_high/frame_000214.jpg", "images/droid_oxe/Remove_the_gray_towel_from_the_carrier_bag_and_put_it_on_the_handle_of_the_couch/episode_0/cam_left_wrist/frame_000214.jpg", "images/droid_oxe/Remove_the_gray_towel_from_the_carrier_bag_and_put_it_on_the_handle_of_the_couch/episode_0/cam_right_wrist/frame_000214.jpg", "images/droid_oxe/Remove_the_gray_towel_from_the_carrier_bag_and_put_it_on_the_handle_of_the_couch/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Remove_the_gray_towel_from_the_carrier_bag_and_put_it_on_the_handle_of_the_couch/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Remove_the_gray_towel_from_the_carrier_bag_and_put_it_on_the_handle_of_the_couch/episode_0/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the gray towel from the carrier bag and put it on the handle of the couch`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the gray towel from the carrier bag and put it on the handle of the couch`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -64.3, "hop_absolute_value": 64.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_orange_ring_on_the_object_episode_0-train-bf_000135-af_000105-minus-0000106", "task": "Put the orange ring on the object", "image": [ "images/droid_oxe/Put_the_orange_ring_on_the_object/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_orange_ring_on_the_object/episode_0/cam_high/frame_000248.jpg", "images/droid_oxe/Put_the_orange_ring_on_the_object/episode_0/cam_high/frame_000135.jpg", "images/droid_oxe/Put_the_orange_ring_on_the_object/episode_0/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Put_the_orange_ring_on_the_object/episode_0/cam_right_wrist/frame_000135.jpg", "images/droid_oxe/Put_the_orange_ring_on_the_object/episode_0/cam_high/frame_000105.jpg", "images/droid_oxe/Put_the_orange_ring_on_the_object/episode_0/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Put_the_orange_ring_on_the_object/episode_0/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the orange ring on the object`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the orange ring on the object`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -22.2, "hop_absolute_value": 22.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_marker_from_the_mug_cup_and_place_it_on_the_table_episode_11-train-bf_000345-af_000240-minus-0000212", "task": "Remove the marker from the mug cup and place it on the table", "image": [ "images/droid_oxe/Remove_the_marker_from_the_mug_cup_and_place_it_on_the_table/episode_11/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_marker_from_the_mug_cup_and_place_it_on_the_table/episode_11/cam_high/frame_000444.jpg", "images/droid_oxe/Remove_the_marker_from_the_mug_cup_and_place_it_on_the_table/episode_11/cam_high/frame_000345.jpg", "images/droid_oxe/Remove_the_marker_from_the_mug_cup_and_place_it_on_the_table/episode_11/cam_left_wrist/frame_000345.jpg", "images/droid_oxe/Remove_the_marker_from_the_mug_cup_and_place_it_on_the_table/episode_11/cam_right_wrist/frame_000345.jpg", "images/droid_oxe/Remove_the_marker_from_the_mug_cup_and_place_it_on_the_table/episode_11/cam_high/frame_000240.jpg", "images/droid_oxe/Remove_the_marker_from_the_mug_cup_and_place_it_on_the_table/episode_11/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Remove_the_marker_from_the_mug_cup_and_place_it_on_the_table/episode_11/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the marker from the mug cup and place it on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the marker from the mug cup and place it on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -30.4, "hop_absolute_value": 30.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_cloth_and_put_it_in_the_box_episode_0-train-bf_000270-af_000135-minus-0000173", "task": "Pick up the cloth and put it in the box", "image": [ "images/droid_oxe/Pick_up_the_cloth_and_put_it_in_the_box/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_cloth_and_put_it_in_the_box/episode_0/cam_high/frame_000424.jpg", "images/droid_oxe/Pick_up_the_cloth_and_put_it_in_the_box/episode_0/cam_high/frame_000270.jpg", "images/droid_oxe/Pick_up_the_cloth_and_put_it_in_the_box/episode_0/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Pick_up_the_cloth_and_put_it_in_the_box/episode_0/cam_right_wrist/frame_000270.jpg", "images/droid_oxe/Pick_up_the_cloth_and_put_it_in_the_box/episode_0/cam_high/frame_000135.jpg", "images/droid_oxe/Pick_up_the_cloth_and_put_it_in_the_box/episode_0/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Pick_up_the_cloth_and_put_it_in_the_box/episode_0/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the cloth and put it in the box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the cloth and put it in the box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Open_the_right_door_of_the_cabinet_episode_6-train-bf_000165-af_000120-minus-0000148", "task": "Open the right door of the cabinet", "image": [ "images/droid_oxe/Open_the_right_door_of_the_cabinet/episode_3/cam_high/frame_000000.jpg", "images/droid_oxe/Open_the_right_door_of_the_cabinet/episode_3/cam_high/frame_000155.jpg", "images/droid_oxe/Open_the_right_door_of_the_cabinet/episode_6/cam_high/frame_000165.jpg", "images/droid_oxe/Open_the_right_door_of_the_cabinet/episode_6/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Open_the_right_door_of_the_cabinet/episode_6/cam_right_wrist/frame_000165.jpg", "images/droid_oxe/Open_the_right_door_of_the_cabinet/episode_6/cam_high/frame_000120.jpg", "images/droid_oxe/Open_the_right_door_of_the_cabinet/episode_6/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Open_the_right_door_of_the_cabinet/episode_6/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the right door of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the right door of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -27.3, "hop_absolute_value": 27.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_blue_towel_to_wipe_the_counter_top_episode_0-train-bf_000165-af_000015-minus-0000032", "task": "Use the blue towel to wipe the counter top", "image": [ "images/droid_oxe/Use_the_blue_towel_to_wipe_the_counter_top/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_blue_towel_to_wipe_the_counter_top/episode_0/cam_high/frame_000315.jpg", "images/droid_oxe/Use_the_blue_towel_to_wipe_the_counter_top/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Use_the_blue_towel_to_wipe_the_counter_top/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Use_the_blue_towel_to_wipe_the_counter_top/episode_0/cam_right_wrist/frame_000165.jpg", "images/droid_oxe/Use_the_blue_towel_to_wipe_the_counter_top/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Use_the_blue_towel_to_wipe_the_counter_top/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Use_the_blue_towel_to_wipe_the_counter_top/episode_0/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the blue towel to wipe the counter top`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the blue towel to wipe the counter top`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -90.9, "hop_absolute_value": 90.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Wipe_the_white_bowl_with_the_white_towel_and_then_put_the_bowl_in_the_bigger_white_bowl_episode_0-train-bf_000285-af_000270-minus-0000268", "task": "Wipe the white bowl with the white towel and then put the bowl in the bigger white bowl", "image": [ "images/droid_oxe/Wipe_the_white_bowl_with_the_white_towel_and_then_put_the_bowl_in_the_bigger_white_bowl/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Wipe_the_white_bowl_with_the_white_towel_and_then_put_the_bowl_in_the_bigger_white_bowl/episode_0/cam_high/frame_000526.jpg", "images/droid_oxe/Wipe_the_white_bowl_with_the_white_towel_and_then_put_the_bowl_in_the_bigger_white_bowl/episode_0/cam_high/frame_000285.jpg", "images/droid_oxe/Wipe_the_white_bowl_with_the_white_towel_and_then_put_the_bowl_in_the_bigger_white_bowl/episode_0/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/Wipe_the_white_bowl_with_the_white_towel_and_then_put_the_bowl_in_the_bigger_white_bowl/episode_0/cam_right_wrist/frame_000285.jpg", "images/droid_oxe/Wipe_the_white_bowl_with_the_white_towel_and_then_put_the_bowl_in_the_bigger_white_bowl/episode_0/cam_high/frame_000270.jpg", "images/droid_oxe/Wipe_the_white_bowl_with_the_white_towel_and_then_put_the_bowl_in_the_bigger_white_bowl/episode_0/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Wipe_the_white_bowl_with_the_white_towel_and_then_put_the_bowl_in_the_bigger_white_bowl/episode_0/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wipe the white bowl with the white towel and then put the bowl in the bigger white bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wipe the white bowl with the white towel and then put the bowl in the bigger white bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.3, "hop_absolute_value": 5.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Uncover_the_small_pot_on_the_stove_episode_0-train-bf_000090-af_000000-minus-0000005", "task": "Uncover the small pot on the stove", "image": [ "images/droid_oxe/Uncover_the_small_pot_on_the_stove/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Uncover_the_small_pot_on_the_stove/episode_0/cam_high/frame_000233.jpg", "images/droid_oxe/Uncover_the_small_pot_on_the_stove/episode_0/cam_high/frame_000090.jpg", "images/droid_oxe/Uncover_the_small_pot_on_the_stove/episode_0/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Uncover_the_small_pot_on_the_stove/episode_0/cam_right_wrist/frame_000090.jpg", "images/droid_oxe/Uncover_the_small_pot_on_the_stove/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Uncover_the_small_pot_on_the_stove/episode_0/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Uncover_the_small_pot_on_the_stove/episode_0/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Uncover the small pot on the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Uncover the small pot on the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_fries_and_put_them_in_the_orange_cup_then_pour_them_into_a_pot_episode_0-train-bf_000270-af_000045-minus-0000080", "task": "Pick up the fries and put them in the orange cup then pour them into a pot", "image": [ "images/droid_oxe/Pick_up_the_fries_and_put_them_in_the_orange_cup_then_pour_them_into_a_pot/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_fries_and_put_them_in_the_orange_cup_then_pour_them_into_a_pot/episode_0/cam_high/frame_000622.jpg", "images/droid_oxe/Pick_up_the_fries_and_put_them_in_the_orange_cup_then_pour_them_into_a_pot/episode_0/cam_high/frame_000270.jpg", "images/droid_oxe/Pick_up_the_fries_and_put_them_in_the_orange_cup_then_pour_them_into_a_pot/episode_0/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Pick_up_the_fries_and_put_them_in_the_orange_cup_then_pour_them_into_a_pot/episode_0/cam_right_wrist/frame_000270.jpg", "images/droid_oxe/Pick_up_the_fries_and_put_them_in_the_orange_cup_then_pour_them_into_a_pot/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Pick_up_the_fries_and_put_them_in_the_orange_cup_then_pour_them_into_a_pot/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Pick_up_the_fries_and_put_them_in_the_orange_cup_then_pour_them_into_a_pot/episode_0/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the fries and put them in the orange cup then pour them into a pot`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the fries and put them in the orange cup then pour them into a pot`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_rope_from_the_black_object_and_put_it_on_the_table_episode_1-train-bf_000195-af_000075-minus-0000066", "task": "Remove the rope from the black object and put it on the table", "image": [ "images/droid_oxe/Remove_the_rope_from_the_black_object_and_put_it_on_the_table/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_rope_from_the_black_object_and_put_it_on_the_table/episode_0/cam_high/frame_000288.jpg", "images/droid_oxe/Remove_the_rope_from_the_black_object_and_put_it_on_the_table/episode_1/cam_high/frame_000195.jpg", "images/droid_oxe/Remove_the_rope_from_the_black_object_and_put_it_on_the_table/episode_1/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Remove_the_rope_from_the_black_object_and_put_it_on_the_table/episode_1/cam_right_wrist/frame_000195.jpg", "images/droid_oxe/Remove_the_rope_from_the_black_object_and_put_it_on_the_table/episode_1/cam_high/frame_000075.jpg", "images/droid_oxe/Remove_the_rope_from_the_black_object_and_put_it_on_the_table/episode_1/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Remove_the_rope_from_the_black_object_and_put_it_on_the_table/episode_1/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the rope from the black object and put it on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the rope from the black object and put it on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -61.5, "hop_absolute_value": 61.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Take_the_spoon_in_the_cup_and_mix_the_contents_in_the_silver_bowl_then_put_the_spoon_back_in_the_cup_and_move_it_next_to_episode_0-train-bf_000690-af_000090-minus-0000073", "task": "Take the spoon in the cup and mix the contents in the silver bowl then put the spoon back in the cup and move it next to", "image": [ "images/droid_oxe/Take_the_spoon_in_the_cup_and_mix_the_contents_in_the_silver_bowl_then_put_the_spoon_back_in_the_cup_and_move_it_next_to/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Take_the_spoon_in_the_cup_and_mix_the_contents_in_the_silver_bowl_then_put_the_spoon_back_in_the_cup_and_move_it_next_to/episode_0/cam_high/frame_000739.jpg", "images/droid_oxe/Take_the_spoon_in_the_cup_and_mix_the_contents_in_the_silver_bowl_then_put_the_spoon_back_in_the_cup_and_move_it_next_to/episode_0/cam_high/frame_000690.jpg", "images/droid_oxe/Take_the_spoon_in_the_cup_and_mix_the_contents_in_the_silver_bowl_then_put_the_spoon_back_in_the_cup_and_move_it_next_to/episode_0/cam_left_wrist/frame_000690.jpg", "images/droid_oxe/Take_the_spoon_in_the_cup_and_mix_the_contents_in_the_silver_bowl_then_put_the_spoon_back_in_the_cup_and_move_it_next_to/episode_0/cam_right_wrist/frame_000690.jpg", "images/droid_oxe/Take_the_spoon_in_the_cup_and_mix_the_contents_in_the_silver_bowl_then_put_the_spoon_back_in_the_cup_and_move_it_next_to/episode_0/cam_high/frame_000090.jpg", "images/droid_oxe/Take_the_spoon_in_the_cup_and_mix_the_contents_in_the_silver_bowl_then_put_the_spoon_back_in_the_cup_and_move_it_next_to/episode_0/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Take_the_spoon_in_the_cup_and_mix_the_contents_in_the_silver_bowl_then_put_the_spoon_back_in_the_cup_and_move_it_next_to/episode_0/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the spoon in the cup and mix the contents in the silver bowl then put the spoon back in the cup and move it next to`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the spoon in the cup and mix the contents in the silver bowl then put the spoon back in the cup and move it next to`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -87.0, "hop_absolute_value": 87.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_wallet_and_put_it_in_the_bag_episode_1-train-bf_000315-af_000105-minus-0000830", "task": "Pick up the wallet and put it in the bag", "image": [ "images/droid_oxe/Pick_up_the_wallet_and_put_it_in_the_bag/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_wallet_and_put_it_in_the_bag/episode_1/cam_high/frame_000452.jpg", "images/droid_oxe/Pick_up_the_wallet_and_put_it_in_the_bag/episode_1/cam_high/frame_000315.jpg", "images/droid_oxe/Pick_up_the_wallet_and_put_it_in_the_bag/episode_1/cam_left_wrist/frame_000315.jpg", "images/droid_oxe/Pick_up_the_wallet_and_put_it_in_the_bag/episode_1/cam_right_wrist/frame_000315.jpg", "images/droid_oxe/Pick_up_the_wallet_and_put_it_in_the_bag/episode_1/cam_high/frame_000105.jpg", "images/droid_oxe/Pick_up_the_wallet_and_put_it_in_the_bag/episode_1/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Pick_up_the_wallet_and_put_it_in_the_bag/episode_1/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the wallet and put it in the bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the wallet and put it in the bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Stack_the_rings_according_to_color_onto_the_wooden_board_on_the_right_episode_0-train-bf_001725-af_000135-minus-0000021", "task": "Stack the rings according to color onto the wooden board on the right", "image": [ "images/droid_oxe/Stack_the_rings_according_to_color_onto_the_wooden_board_on_the_right/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Stack_the_rings_according_to_color_onto_the_wooden_board_on_the_right/episode_0/cam_high/frame_003117.jpg", "images/droid_oxe/Stack_the_rings_according_to_color_onto_the_wooden_board_on_the_right/episode_0/cam_high/frame_001725.jpg", "images/droid_oxe/Stack_the_rings_according_to_color_onto_the_wooden_board_on_the_right/episode_0/cam_left_wrist/frame_001725.jpg", "images/droid_oxe/Stack_the_rings_according_to_color_onto_the_wooden_board_on_the_right/episode_0/cam_right_wrist/frame_001725.jpg", "images/droid_oxe/Stack_the_rings_according_to_color_onto_the_wooden_board_on_the_right/episode_0/cam_high/frame_000135.jpg", "images/droid_oxe/Stack_the_rings_according_to_color_onto_the_wooden_board_on_the_right/episode_0/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Stack_the_rings_according_to_color_onto_the_wooden_board_on_the_right/episode_0/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Stack the rings according to color onto the wooden board on the right`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Stack the rings according to color onto the wooden board on the right`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -92.2, "hop_absolute_value": 92.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Place_the_white_rope_on_the_curved_black_board_episode_0-train-bf_000480-af_000090-minus-0000089", "task": "Place the white rope on the curved black board", "image": [ "images/droid_oxe/Place_the_white_rope_on_the_curved_black_board/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Place_the_white_rope_on_the_curved_black_board/episode_0/cam_high/frame_000930.jpg", "images/droid_oxe/Place_the_white_rope_on_the_curved_black_board/episode_0/cam_high/frame_000480.jpg", "images/droid_oxe/Place_the_white_rope_on_the_curved_black_board/episode_0/cam_left_wrist/frame_000480.jpg", "images/droid_oxe/Place_the_white_rope_on_the_curved_black_board/episode_0/cam_right_wrist/frame_000480.jpg", "images/droid_oxe/Place_the_white_rope_on_the_curved_black_board/episode_0/cam_high/frame_000090.jpg", "images/droid_oxe/Place_the_white_rope_on_the_curved_black_board/episode_0/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Place_the_white_rope_on_the_curved_black_board/episode_0/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the white rope on the curved black board`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the white rope on the curved black board`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -81.2, "hop_absolute_value": 81.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_peach_shirt_and_the_white_towel_to_the_right_then_move_the_pink_shirt_to_the_left_and_move_the_turnip_plushie_t_episode_0-train-bf_000300-af_000165-minus-0000190", "task": "Move the peach shirt and the white towel to the right then move the pink shirt to the left and move the turnip plushie t", "image": [ "images/droid_oxe/Move_the_peach_shirt_and_the_white_towel_to_the_right_then_move_the_pink_shirt_to_the_left_and_move_the_turnip_plushie_t/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_peach_shirt_and_the_white_towel_to_the_right_then_move_the_pink_shirt_to_the_left_and_move_the_turnip_plushie_t/episode_0/cam_high/frame_000537.jpg", "images/droid_oxe/Move_the_peach_shirt_and_the_white_towel_to_the_right_then_move_the_pink_shirt_to_the_left_and_move_the_turnip_plushie_t/episode_0/cam_high/frame_000300.jpg", "images/droid_oxe/Move_the_peach_shirt_and_the_white_towel_to_the_right_then_move_the_pink_shirt_to_the_left_and_move_the_turnip_plushie_t/episode_0/cam_left_wrist/frame_000300.jpg", "images/droid_oxe/Move_the_peach_shirt_and_the_white_towel_to_the_right_then_move_the_pink_shirt_to_the_left_and_move_the_turnip_plushie_t/episode_0/cam_right_wrist/frame_000300.jpg", "images/droid_oxe/Move_the_peach_shirt_and_the_white_towel_to_the_right_then_move_the_pink_shirt_to_the_left_and_move_the_turnip_plushie_t/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Move_the_peach_shirt_and_the_white_towel_to_the_right_then_move_the_pink_shirt_to_the_left_and_move_the_turnip_plushie_t/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Move_the_peach_shirt_and_the_white_towel_to_the_right_then_move_the_pink_shirt_to_the_left_and_move_the_turnip_plushie_t/episode_0/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the peach shirt and the white towel to the right then move the pink shirt to the left and move the turnip plushie t`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the peach shirt and the white towel to the right then move the pink shirt to the left and move the turnip plushie t`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -45.0, "hop_absolute_value": 45.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_black_gadget_on_the_circle_shaped_tray_and_the_keys_on_the_square_tray_episode_0-train-bf_000465-af_000240-minus-0000204", "task": "Put the black gadget on the circle shaped tray and the keys on the square tray", "image": [ "images/droid_oxe/Put_the_black_gadget_on_the_circle_shaped_tray_and_the_keys_on_the_square_tray/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_black_gadget_on_the_circle_shaped_tray_and_the_keys_on_the_square_tray/episode_0/cam_high/frame_000572.jpg", "images/droid_oxe/Put_the_black_gadget_on_the_circle_shaped_tray_and_the_keys_on_the_square_tray/episode_0/cam_high/frame_000465.jpg", "images/droid_oxe/Put_the_black_gadget_on_the_circle_shaped_tray_and_the_keys_on_the_square_tray/episode_0/cam_left_wrist/frame_000465.jpg", "images/droid_oxe/Put_the_black_gadget_on_the_circle_shaped_tray_and_the_keys_on_the_square_tray/episode_0/cam_right_wrist/frame_000465.jpg", "images/droid_oxe/Put_the_black_gadget_on_the_circle_shaped_tray_and_the_keys_on_the_square_tray/episode_0/cam_high/frame_000240.jpg", "images/droid_oxe/Put_the_black_gadget_on_the_circle_shaped_tray_and_the_keys_on_the_square_tray/episode_0/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Put_the_black_gadget_on_the_circle_shaped_tray_and_the_keys_on_the_square_tray/episode_0/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the black gadget on the circle shaped tray and the keys on the square tray`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the black gadget on the circle shaped tray and the keys on the square tray`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -48.4, "hop_absolute_value": 48.4, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Press_the_button_on_the_white_lid_then_put_the_white_lid_on_the_table_episode_0-train-bf_000045-af_000030-minus-0000674", "task": "Press the button on the white lid then put the white lid on the table", "image": [ "images/droid_oxe/Press_the_button_on_the_white_lid_then_put_the_white_lid_on_the_table/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Press_the_button_on_the_white_lid_then_put_the_white_lid_on_the_table/episode_0/cam_high/frame_000444.jpg", "images/droid_oxe/Press_the_button_on_the_white_lid_then_put_the_white_lid_on_the_table/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Press_the_button_on_the_white_lid_then_put_the_white_lid_on_the_table/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Press_the_button_on_the_white_lid_then_put_the_white_lid_on_the_table/episode_0/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Press_the_button_on_the_white_lid_then_put_the_white_lid_on_the_table/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Press_the_button_on_the_white_lid_then_put_the_white_lid_on_the_table/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Press_the_button_on_the_white_lid_then_put_the_white_lid_on_the_table/episode_0/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Press the button on the white lid then put the white lid on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Press the button on the white lid then put the white lid on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_brown_clothing_on_the_bed_in_the_laundry_hamper_episode_0-train-bf_000225-af_000060-minus-0000088", "task": "Put the brown clothing on the bed in the laundry hamper", "image": [ "images/droid_oxe/Put_the_brown_clothing_on_the_bed_in_the_laundry_hamper/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_brown_clothing_on_the_bed_in_the_laundry_hamper/episode_0/cam_high/frame_000363.jpg", "images/droid_oxe/Put_the_brown_clothing_on_the_bed_in_the_laundry_hamper/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Put_the_brown_clothing_on_the_bed_in_the_laundry_hamper/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_brown_clothing_on_the_bed_in_the_laundry_hamper/episode_0/cam_right_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_brown_clothing_on_the_bed_in_the_laundry_hamper/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Put_the_brown_clothing_on_the_bed_in_the_laundry_hamper/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Put_the_brown_clothing_on_the_bed_in_the_laundry_hamper/episode_0/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the brown clothing on the bed in the laundry hamper`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the brown clothing on the bed in the laundry hamper`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.3, "hop_absolute_value": 73.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_rod_on_the_couch_arm_rest_episode_0-train-bf_000225-af_000150-minus-0000544", "task": "Put the rod on the couch arm rest", "image": [ "images/droid_oxe/Put_the_rod_on_the_couch_arm_rest/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_rod_on_the_couch_arm_rest/episode_0/cam_high/frame_000360.jpg", "images/droid_oxe/Put_the_rod_on_the_couch_arm_rest/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Put_the_rod_on_the_couch_arm_rest/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_rod_on_the_couch_arm_rest/episode_0/cam_right_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_rod_on_the_couch_arm_rest/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Put_the_rod_on_the_couch_arm_rest/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Put_the_rod_on_the_couch_arm_rest/episode_0/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the rod on the couch arm rest`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the rod on the couch arm rest`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_three_water_bottles_to_the_right_episode_0-train-bf_000523-af_000495-minus-0000567", "task": "Move three water bottles to the right", "image": [ "images/droid_oxe/Move_three_water_bottles_to_the_right/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_three_water_bottles_to_the_right/episode_0/cam_high/frame_000523.jpg", "images/droid_oxe/Move_three_water_bottles_to_the_right/episode_0/cam_high/frame_000523.jpg", "images/droid_oxe/Move_three_water_bottles_to_the_right/episode_0/cam_left_wrist/frame_000523.jpg", "images/droid_oxe/Move_three_water_bottles_to_the_right/episode_0/cam_right_wrist/frame_000523.jpg", "images/droid_oxe/Move_three_water_bottles_to_the_right/episode_0/cam_high/frame_000495.jpg", "images/droid_oxe/Move_three_water_bottles_to_the_right/episode_0/cam_left_wrist/frame_000495.jpg", "images/droid_oxe/Move_three_water_bottles_to_the_right/episode_0/cam_right_wrist/frame_000495.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move three water bottles to the right`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move three water bottles to the right`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.9, "hop_absolute_value": 2.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Fold_the_grey_blanket_once_episode_0-train-bf_000240-af_000210-minus-0000360", "task": "Fold the grey blanket once", "image": [ "images/droid_oxe/Fold_the_grey_blanket_once/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Fold_the_grey_blanket_once/episode_0/cam_high/frame_000294.jpg", "images/droid_oxe/Fold_the_grey_blanket_once/episode_0/cam_high/frame_000240.jpg", "images/droid_oxe/Fold_the_grey_blanket_once/episode_0/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Fold_the_grey_blanket_once/episode_0/cam_right_wrist/frame_000240.jpg", "images/droid_oxe/Fold_the_grey_blanket_once/episode_0/cam_high/frame_000210.jpg", "images/droid_oxe/Fold_the_grey_blanket_once/episode_0/cam_left_wrist/frame_000210.jpg", "images/droid_oxe/Fold_the_grey_blanket_once/episode_0/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the grey blanket once`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the grey blanket once`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_plushy_and_peach_shirt_on_the_right_and_put_the_pink_shirt_on_the_left_episode_0-train-bf_000255-af_000135-minus-0000180", "task": "Put the plushy and peach shirt on the right and put the pink shirt on the left", "image": [ "images/droid_oxe/Put_the_plushy_and_peach_shirt_on_the_right_and_put_the_pink_shirt_on_the_left/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_plushy_and_peach_shirt_on_the_right_and_put_the_pink_shirt_on_the_left/episode_0/cam_high/frame_000433.jpg", "images/droid_oxe/Put_the_plushy_and_peach_shirt_on_the_right_and_put_the_pink_shirt_on_the_left/episode_0/cam_high/frame_000255.jpg", "images/droid_oxe/Put_the_plushy_and_peach_shirt_on_the_right_and_put_the_pink_shirt_on_the_left/episode_0/cam_left_wrist/frame_000255.jpg", "images/droid_oxe/Put_the_plushy_and_peach_shirt_on_the_right_and_put_the_pink_shirt_on_the_left/episode_0/cam_right_wrist/frame_000255.jpg", "images/droid_oxe/Put_the_plushy_and_peach_shirt_on_the_right_and_put_the_pink_shirt_on_the_left/episode_0/cam_high/frame_000135.jpg", "images/droid_oxe/Put_the_plushy_and_peach_shirt_on_the_right_and_put_the_pink_shirt_on_the_left/episode_0/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Put_the_plushy_and_peach_shirt_on_the_right_and_put_the_pink_shirt_on_the_left/episode_0/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the plushy and peach shirt on the right and put the pink shirt on the left`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the plushy and peach shirt on the right and put the pink shirt on the left`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -47.1, "hop_absolute_value": 47.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_green_object_from_the_bowl_episode_3-train-bf_000120-af_000015-minus-0000010", "task": "Remove the green object from the bowl", "image": [ "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_1/cam_high/frame_000182.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_3/cam_high/frame_000120.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_3/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_3/cam_right_wrist/frame_000120.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_3/cam_high/frame_000015.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_3/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Remove_the_green_object_from_the_bowl/episode_3/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the green object from the bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the green object from the bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Take_the_marker_out_of_the_mug_put_it_on_the_table_then_put_it_back_in_the_mug_episode_0-train-bf_000375-af_000285-minus-0000914", "task": "Take the marker out of the mug put it on the table then put it back in the mug", "image": [ "images/droid_oxe/Take_the_marker_out_of_the_mug_put_it_on_the_table_then_put_it_back_in_the_mug/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Take_the_marker_out_of_the_mug_put_it_on_the_table_then_put_it_back_in_the_mug/episode_0/cam_high/frame_000670.jpg", "images/droid_oxe/Take_the_marker_out_of_the_mug_put_it_on_the_table_then_put_it_back_in_the_mug/episode_0/cam_high/frame_000375.jpg", "images/droid_oxe/Take_the_marker_out_of_the_mug_put_it_on_the_table_then_put_it_back_in_the_mug/episode_0/cam_left_wrist/frame_000375.jpg", "images/droid_oxe/Take_the_marker_out_of_the_mug_put_it_on_the_table_then_put_it_back_in_the_mug/episode_0/cam_right_wrist/frame_000375.jpg", "images/droid_oxe/Take_the_marker_out_of_the_mug_put_it_on_the_table_then_put_it_back_in_the_mug/episode_0/cam_high/frame_000285.jpg", "images/droid_oxe/Take_the_marker_out_of_the_mug_put_it_on_the_table_then_put_it_back_in_the_mug/episode_0/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/Take_the_marker_out_of_the_mug_put_it_on_the_table_then_put_it_back_in_the_mug/episode_0/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the marker out of the mug put it on the table then put it back in the mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the marker out of the mug put it on the table then put it back in the mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -24.0, "hop_absolute_value": 24.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_bowl_to_the_right_episode_14-train-bf_000099-af_000060-minus-0000017", "task": "Move the bowl to the right", "image": [ "images/droid_oxe/Move_the_bowl_to_the_right/episode_14/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_bowl_to_the_right/episode_14/cam_high/frame_000099.jpg", "images/droid_oxe/Move_the_bowl_to_the_right/episode_14/cam_high/frame_000099.jpg", "images/droid_oxe/Move_the_bowl_to_the_right/episode_14/cam_left_wrist/frame_000099.jpg", "images/droid_oxe/Move_the_bowl_to_the_right/episode_14/cam_right_wrist/frame_000099.jpg", "images/droid_oxe/Move_the_bowl_to_the_right/episode_14/cam_high/frame_000060.jpg", "images/droid_oxe/Move_the_bowl_to_the_right/episode_14/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Move_the_bowl_to_the_right/episode_14/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the bowl to the right`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the bowl to the right`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Rotate_the_sneaker_90_degrees_clockwise_episode_0-train-bf_000150-af_000015-minus-0000014", "task": "Rotate the sneaker 90 degrees clockwise", "image": [ "images/droid_oxe/Rotate_the_sneaker_90_degrees_clockwise/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Rotate_the_sneaker_90_degrees_clockwise/episode_0/cam_high/frame_000166.jpg", "images/droid_oxe/Rotate_the_sneaker_90_degrees_clockwise/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Rotate_the_sneaker_90_degrees_clockwise/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Rotate_the_sneaker_90_degrees_clockwise/episode_0/cam_right_wrist/frame_000150.jpg", "images/droid_oxe/Rotate_the_sneaker_90_degrees_clockwise/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Rotate_the_sneaker_90_degrees_clockwise/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Rotate_the_sneaker_90_degrees_clockwise/episode_0/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Rotate the sneaker 90 degrees clockwise`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Rotate the sneaker 90 degrees clockwise`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -90.0, "hop_absolute_value": 90.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pour_the_things_in_the_bowl_onto_the_table_and_then_sweep_them_to_the_left_episode_0-train-bf_000315-af_000195-minus-0000220", "task": "Pour the things in the bowl onto the table and then sweep them to the left", "image": [ "images/droid_oxe/Pour_the_things_in_the_bowl_onto_the_table_and_then_sweep_them_to_the_left/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pour_the_things_in_the_bowl_onto_the_table_and_then_sweep_them_to_the_left/episode_0/cam_high/frame_000695.jpg", "images/droid_oxe/Pour_the_things_in_the_bowl_onto_the_table_and_then_sweep_them_to_the_left/episode_0/cam_high/frame_000315.jpg", "images/droid_oxe/Pour_the_things_in_the_bowl_onto_the_table_and_then_sweep_them_to_the_left/episode_0/cam_left_wrist/frame_000315.jpg", "images/droid_oxe/Pour_the_things_in_the_bowl_onto_the_table_and_then_sweep_them_to_the_left/episode_0/cam_right_wrist/frame_000315.jpg", "images/droid_oxe/Pour_the_things_in_the_bowl_onto_the_table_and_then_sweep_them_to_the_left/episode_0/cam_high/frame_000195.jpg", "images/droid_oxe/Pour_the_things_in_the_bowl_onto_the_table_and_then_sweep_them_to_the_left/episode_0/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Pour_the_things_in_the_bowl_onto_the_table_and_then_sweep_them_to_the_left/episode_0/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pour the things in the bowl onto the table and then sweep them to the left`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pour the things in the bowl onto the table and then sweep them to the left`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -38.1, "hop_absolute_value": 38.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Unstack_the_measuring_cups_episode_0-train-bf_000450-af_000375-minus-0000880", "task": "Unstack the measuring cups", "image": [ "images/droid_oxe/Unstack_the_measuring_cups/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Unstack_the_measuring_cups/episode_1/cam_high/frame_000491.jpg", "images/droid_oxe/Unstack_the_measuring_cups/episode_0/cam_high/frame_000450.jpg", "images/droid_oxe/Unstack_the_measuring_cups/episode_0/cam_left_wrist/frame_000450.jpg", "images/droid_oxe/Unstack_the_measuring_cups/episode_0/cam_right_wrist/frame_000450.jpg", "images/droid_oxe/Unstack_the_measuring_cups/episode_0/cam_high/frame_000375.jpg", "images/droid_oxe/Unstack_the_measuring_cups/episode_0/cam_left_wrist/frame_000375.jpg", "images/droid_oxe/Unstack_the_measuring_cups/episode_0/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unstack the measuring cups`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unstack the measuring cups`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Open_the_cabinet_door_on_the_far_right_episode_2-train-bf_000090-af_000015-minus-0000022", "task": "Open the cabinet door on the far right", "image": [ "images/droid_oxe/Open_the_cabinet_door_on_the_far_right/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Open_the_cabinet_door_on_the_far_right/episode_0/cam_high/frame_000219.jpg", "images/droid_oxe/Open_the_cabinet_door_on_the_far_right/episode_2/cam_high/frame_000090.jpg", "images/droid_oxe/Open_the_cabinet_door_on_the_far_right/episode_2/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Open_the_cabinet_door_on_the_far_right/episode_2/cam_right_wrist/frame_000090.jpg", "images/droid_oxe/Open_the_cabinet_door_on_the_far_right/episode_2/cam_high/frame_000015.jpg", "images/droid_oxe/Open_the_cabinet_door_on_the_far_right/episode_2/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Open_the_cabinet_door_on_the_far_right/episode_2/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the cabinet door on the far right`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the cabinet door on the far right`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_green_cloth_from_the_drawer_and_place_it_on_the_bed_episode_0-train-bf_000195-af_000000-minus-0000011", "task": "Remove the green cloth from the drawer and place it on the bed", "image": [ "images/droid_oxe/Remove_the_green_cloth_from_the_drawer_and_place_it_on_the_bed/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_green_cloth_from_the_drawer_and_place_it_on_the_bed/episode_0/cam_high/frame_000239.jpg", "images/droid_oxe/Remove_the_green_cloth_from_the_drawer_and_place_it_on_the_bed/episode_0/cam_high/frame_000195.jpg", "images/droid_oxe/Remove_the_green_cloth_from_the_drawer_and_place_it_on_the_bed/episode_0/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Remove_the_green_cloth_from_the_drawer_and_place_it_on_the_bed/episode_0/cam_right_wrist/frame_000195.jpg", "images/droid_oxe/Remove_the_green_cloth_from_the_drawer_and_place_it_on_the_bed/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_green_cloth_from_the_drawer_and_place_it_on_the_bed/episode_0/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Remove_the_green_cloth_from_the_drawer_and_place_it_on_the_bed/episode_0/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the green cloth from the drawer and place it on the bed`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the green cloth from the drawer and place it on the bed`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Cover_the_fruits_plush_toy_with_the_blue_cloth_episode_0-train-bf_000465-af_000225-minus-0000882", "task": "Cover the fruits plush toy with the blue cloth", "image": [ "images/droid_oxe/Cover_the_fruits_plush_toy_with_the_blue_cloth/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Cover_the_fruits_plush_toy_with_the_blue_cloth/episode_0/cam_high/frame_000505.jpg", "images/droid_oxe/Cover_the_fruits_plush_toy_with_the_blue_cloth/episode_0/cam_high/frame_000465.jpg", "images/droid_oxe/Cover_the_fruits_plush_toy_with_the_blue_cloth/episode_0/cam_left_wrist/frame_000465.jpg", "images/droid_oxe/Cover_the_fruits_plush_toy_with_the_blue_cloth/episode_0/cam_right_wrist/frame_000465.jpg", "images/droid_oxe/Cover_the_fruits_plush_toy_with_the_blue_cloth/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Cover_the_fruits_plush_toy_with_the_blue_cloth/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Cover_the_fruits_plush_toy_with_the_blue_cloth/episode_0/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Cover the fruits plush toy with the blue cloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Cover the fruits plush toy with the blue cloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -51.6, "hop_absolute_value": 51.6, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pour_the_contents_of_the_orange_bowl_into_the_white_cup_episode_0-train-bf_000360-af_000345-minus-0000549", "task": "Pour the contents of the orange bowl into the white cup", "image": [ "images/droid_oxe/Pour_the_contents_of_the_orange_bowl_into_the_white_cup/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pour_the_contents_of_the_orange_bowl_into_the_white_cup/episode_0/cam_high/frame_000447.jpg", "images/droid_oxe/Pour_the_contents_of_the_orange_bowl_into_the_white_cup/episode_0/cam_high/frame_000360.jpg", "images/droid_oxe/Pour_the_contents_of_the_orange_bowl_into_the_white_cup/episode_0/cam_left_wrist/frame_000360.jpg", "images/droid_oxe/Pour_the_contents_of_the_orange_bowl_into_the_white_cup/episode_0/cam_right_wrist/frame_000360.jpg", "images/droid_oxe/Pour_the_contents_of_the_orange_bowl_into_the_white_cup/episode_0/cam_high/frame_000345.jpg", "images/droid_oxe/Pour_the_contents_of_the_orange_bowl_into_the_white_cup/episode_0/cam_left_wrist/frame_000345.jpg", "images/droid_oxe/Pour_the_contents_of_the_orange_bowl_into_the_white_cup/episode_0/cam_right_wrist/frame_000345.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pour the contents of the orange bowl into the white cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pour the contents of the orange bowl into the white cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.2, "hop_absolute_value": 4.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_all_the_clothes_from_the_bowl_and_put_them_on_the_table_episode_0-train-bf_000255-af_000060-minus-0000069", "task": "Remove all the clothes from the bowl and put them on the table", "image": [ "images/droid_oxe/Remove_all_the_clothes_from_the_bowl_and_put_them_on_the_table/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_all_the_clothes_from_the_bowl_and_put_them_on_the_table/episode_0/cam_high/frame_000409.jpg", "images/droid_oxe/Remove_all_the_clothes_from_the_bowl_and_put_them_on_the_table/episode_0/cam_high/frame_000255.jpg", "images/droid_oxe/Remove_all_the_clothes_from_the_bowl_and_put_them_on_the_table/episode_0/cam_left_wrist/frame_000255.jpg", "images/droid_oxe/Remove_all_the_clothes_from_the_bowl_and_put_them_on_the_table/episode_0/cam_right_wrist/frame_000255.jpg", "images/droid_oxe/Remove_all_the_clothes_from_the_bowl_and_put_them_on_the_table/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Remove_all_the_clothes_from_the_bowl_and_put_them_on_the_table/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Remove_all_the_clothes_from_the_bowl_and_put_them_on_the_table/episode_0/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove all the clothes from the bowl and put them on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove all the clothes from the bowl and put them on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -76.5, "hop_absolute_value": 76.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_grey_pot_on_the_top_left_plate_of_the_stove_episode_1-train-bf_000240-af_000150-minus-0000184", "task": "Put the grey pot on the top left plate of the stove", "image": [ "images/droid_oxe/Put_the_grey_pot_on_the_top_left_plate_of_the_stove/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_grey_pot_on_the_top_left_plate_of_the_stove/episode_1/cam_high/frame_000433.jpg", "images/droid_oxe/Put_the_grey_pot_on_the_top_left_plate_of_the_stove/episode_1/cam_high/frame_000240.jpg", "images/droid_oxe/Put_the_grey_pot_on_the_top_left_plate_of_the_stove/episode_1/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Put_the_grey_pot_on_the_top_left_plate_of_the_stove/episode_1/cam_right_wrist/frame_000240.jpg", "images/droid_oxe/Put_the_grey_pot_on_the_top_left_plate_of_the_stove/episode_1/cam_high/frame_000150.jpg", "images/droid_oxe/Put_the_grey_pot_on_the_top_left_plate_of_the_stove/episode_1/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Put_the_grey_pot_on_the_top_left_plate_of_the_stove/episode_1/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the grey pot on the top left plate of the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the grey pot on the top left plate of the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -37.5, "hop_absolute_value": 37.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_move_the_plate_from_the_shelf_to_the_counter_episode_0-train-bf_000285-af_000000-minus-0000021", "task": "move the plate from the shelf to the counter", "image": [ "images/droid_oxe/move_the_plate_from_the_shelf_to_the_counter/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/move_the_plate_from_the_shelf_to_the_counter/episode_0/cam_high/frame_000302.jpg", "images/droid_oxe/move_the_plate_from_the_shelf_to_the_counter/episode_0/cam_high/frame_000285.jpg", "images/droid_oxe/move_the_plate_from_the_shelf_to_the_counter/episode_0/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/move_the_plate_from_the_shelf_to_the_counter/episode_0/cam_right_wrist/frame_000285.jpg", "images/droid_oxe/move_the_plate_from_the_shelf_to_the_counter/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/move_the_plate_from_the_shelf_to_the_counter/episode_0/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/move_the_plate_from_the_shelf_to_the_counter/episode_0/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`move the plate from the shelf to the counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `move the plate from the shelf to the counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Open_the_top_right_cabinet_door_episode_3-train-bf_000045-af_000000-minus-0000001", "task": "Open the top right cabinet door", "image": [ "images/droid_oxe/Open_the_top_right_cabinet_door/episode_9/cam_high/frame_000000.jpg", "images/droid_oxe/Open_the_top_right_cabinet_door/episode_9/cam_high/frame_000579.jpg", "images/droid_oxe/Open_the_top_right_cabinet_door/episode_3/cam_high/frame_000045.jpg", "images/droid_oxe/Open_the_top_right_cabinet_door/episode_3/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Open_the_top_right_cabinet_door/episode_3/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Open_the_top_right_cabinet_door/episode_3/cam_high/frame_000000.jpg", "images/droid_oxe/Open_the_top_right_cabinet_door/episode_3/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Open_the_top_right_cabinet_door/episode_3/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the top right cabinet door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the top right cabinet door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_silver_bowl_to_the_right_and_move_the_pan_toward_the_edge_of_the_stove_episode_0-train-bf_000270-af_000105-minus-0000114", "task": "Move the silver bowl to the right and move the pan toward the edge of the stove", "image": [ "images/droid_oxe/Move_the_silver_bowl_to_the_right_and_move_the_pan_toward_the_edge_of_the_stove/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_silver_bowl_to_the_right_and_move_the_pan_toward_the_edge_of_the_stove/episode_0/cam_high/frame_000378.jpg", "images/droid_oxe/Move_the_silver_bowl_to_the_right_and_move_the_pan_toward_the_edge_of_the_stove/episode_0/cam_high/frame_000270.jpg", "images/droid_oxe/Move_the_silver_bowl_to_the_right_and_move_the_pan_toward_the_edge_of_the_stove/episode_0/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Move_the_silver_bowl_to_the_right_and_move_the_pan_toward_the_edge_of_the_stove/episode_0/cam_right_wrist/frame_000270.jpg", "images/droid_oxe/Move_the_silver_bowl_to_the_right_and_move_the_pan_toward_the_edge_of_the_stove/episode_0/cam_high/frame_000105.jpg", "images/droid_oxe/Move_the_silver_bowl_to_the_right_and_move_the_pan_toward_the_edge_of_the_stove/episode_0/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Move_the_silver_bowl_to_the_right_and_move_the_pan_toward_the_edge_of_the_stove/episode_0/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the silver bowl to the right and move the pan toward the edge of the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the silver bowl to the right and move the pan toward the edge of the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -61.1, "hop_absolute_value": 61.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_blue_cup_in_the_yellow_bowl_then_put_the_yellow_bowl_on_the_gray_plate_episode_0-train-bf_000225-af_000105-minus-0000147", "task": "Put the blue cup in the yellow bowl then put the yellow bowl on the gray plate", "image": [ "images/droid_oxe/Put_the_blue_cup_in_the_yellow_bowl_then_put_the_yellow_bowl_on_the_gray_plate/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_blue_cup_in_the_yellow_bowl_then_put_the_yellow_bowl_on_the_gray_plate/episode_0/cam_high/frame_000433.jpg", "images/droid_oxe/Put_the_blue_cup_in_the_yellow_bowl_then_put_the_yellow_bowl_on_the_gray_plate/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Put_the_blue_cup_in_the_yellow_bowl_then_put_the_yellow_bowl_on_the_gray_plate/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_blue_cup_in_the_yellow_bowl_then_put_the_yellow_bowl_on_the_gray_plate/episode_0/cam_right_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_blue_cup_in_the_yellow_bowl_then_put_the_yellow_bowl_on_the_gray_plate/episode_0/cam_high/frame_000105.jpg", "images/droid_oxe/Put_the_blue_cup_in_the_yellow_bowl_then_put_the_yellow_bowl_on_the_gray_plate/episode_0/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Put_the_blue_cup_in_the_yellow_bowl_then_put_the_yellow_bowl_on_the_gray_plate/episode_0/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the blue cup in the yellow bowl then put the yellow bowl on the gray plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the blue cup in the yellow bowl then put the yellow bowl on the gray plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -53.3, "hop_absolute_value": 53.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_cob_of_maize_in_the_pot_episode_0-train-bf_000285-af_000165-minus-0000150", "task": "Put the cob of maize in the pot", "image": [ "images/droid_oxe/Put_the_cob_of_maize_in_the_pot/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_cob_of_maize_in_the_pot/episode_0/cam_high/frame_000315.jpg", "images/droid_oxe/Put_the_cob_of_maize_in_the_pot/episode_0/cam_high/frame_000285.jpg", "images/droid_oxe/Put_the_cob_of_maize_in_the_pot/episode_0/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/Put_the_cob_of_maize_in_the_pot/episode_0/cam_right_wrist/frame_000285.jpg", "images/droid_oxe/Put_the_cob_of_maize_in_the_pot/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Put_the_cob_of_maize_in_the_pot/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Put_the_cob_of_maize_in_the_pot/episode_0/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the cob of maize in the pot`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the cob of maize in the pot`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -42.1, "hop_absolute_value": 42.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pour_the_contents_in_the_basket_into_the_coffee_cup_episode_0-train-bf_000345-af_000000-minus-0000030", "task": "Pour the contents in the basket into the coffee cup", "image": [ "images/droid_oxe/Pour_the_contents_in_the_basket_into_the_coffee_cup/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pour_the_contents_in_the_basket_into_the_coffee_cup/episode_0/cam_high/frame_000366.jpg", "images/droid_oxe/Pour_the_contents_in_the_basket_into_the_coffee_cup/episode_0/cam_high/frame_000345.jpg", "images/droid_oxe/Pour_the_contents_in_the_basket_into_the_coffee_cup/episode_0/cam_left_wrist/frame_000345.jpg", "images/droid_oxe/Pour_the_contents_in_the_basket_into_the_coffee_cup/episode_0/cam_right_wrist/frame_000345.jpg", "images/droid_oxe/Pour_the_contents_in_the_basket_into_the_coffee_cup/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pour_the_contents_in_the_basket_into_the_coffee_cup/episode_0/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Pour_the_contents_in_the_basket_into_the_coffee_cup/episode_0/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pour the contents in the basket into the coffee cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pour the contents in the basket into the coffee cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_orange_chip_into_the_connect_four_game_episode_0-train-bf_000315-af_000180-minus-0000178", "task": "Put the orange chip into the connect four game", "image": [ "images/droid_oxe/Put_the_orange_chip_into_the_connect_four_game/episode_2/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_orange_chip_into_the_connect_four_game/episode_2/cam_high/frame_000366.jpg", "images/droid_oxe/Put_the_orange_chip_into_the_connect_four_game/episode_0/cam_high/frame_000315.jpg", "images/droid_oxe/Put_the_orange_chip_into_the_connect_four_game/episode_0/cam_left_wrist/frame_000315.jpg", "images/droid_oxe/Put_the_orange_chip_into_the_connect_four_game/episode_0/cam_right_wrist/frame_000315.jpg", "images/droid_oxe/Put_the_orange_chip_into_the_connect_four_game/episode_0/cam_high/frame_000180.jpg", "images/droid_oxe/Put_the_orange_chip_into_the_connect_four_game/episode_0/cam_left_wrist/frame_000180.jpg", "images/droid_oxe/Put_the_orange_chip_into_the_connect_four_game/episode_0/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the orange chip into the connect four game`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the orange chip into the connect four game`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -42.9, "hop_absolute_value": 42.9, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Push_the_faucet_to_the_left_use_the_sponge_in_the_sink_to_wipe_the_sink_episode_0-train-bf_000210-af_000180-minus-0000903", "task": "Push the faucet to the left use the sponge in the sink to wipe the sink", "image": [ "images/droid_oxe/Push_the_faucet_to_the_left_use_the_sponge_in_the_sink_to_wipe_the_sink/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Push_the_faucet_to_the_left_use_the_sponge_in_the_sink_to_wipe_the_sink/episode_0/cam_high/frame_000789.jpg", "images/droid_oxe/Push_the_faucet_to_the_left_use_the_sponge_in_the_sink_to_wipe_the_sink/episode_0/cam_high/frame_000210.jpg", "images/droid_oxe/Push_the_faucet_to_the_left_use_the_sponge_in_the_sink_to_wipe_the_sink/episode_0/cam_left_wrist/frame_000210.jpg", "images/droid_oxe/Push_the_faucet_to_the_left_use_the_sponge_in_the_sink_to_wipe_the_sink/episode_0/cam_right_wrist/frame_000210.jpg", "images/droid_oxe/Push_the_faucet_to_the_left_use_the_sponge_in_the_sink_to_wipe_the_sink/episode_0/cam_high/frame_000180.jpg", "images/droid_oxe/Push_the_faucet_to_the_left_use_the_sponge_in_the_sink_to_wipe_the_sink/episode_0/cam_left_wrist/frame_000180.jpg", "images/droid_oxe/Push_the_faucet_to_the_left_use_the_sponge_in_the_sink_to_wipe_the_sink/episode_0/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the faucet to the left use the sponge in the sink to wipe the sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the faucet to the left use the sponge in the sink to wipe the sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -14.3, "hop_absolute_value": 14.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_yellow_object_yellow_pencil_and_pink_object_inside_the_open_cabinet_episode_0-train-bf_000435-af_000360-minus-0000270", "task": "Put the yellow object yellow pencil and pink object inside the open cabinet", "image": [ "images/droid_oxe/Put_the_yellow_object_yellow_pencil_and_pink_object_inside_the_open_cabinet/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_yellow_object_yellow_pencil_and_pink_object_inside_the_open_cabinet/episode_0/cam_high/frame_000640.jpg", "images/droid_oxe/Put_the_yellow_object_yellow_pencil_and_pink_object_inside_the_open_cabinet/episode_0/cam_high/frame_000435.jpg", "images/droid_oxe/Put_the_yellow_object_yellow_pencil_and_pink_object_inside_the_open_cabinet/episode_0/cam_left_wrist/frame_000435.jpg", "images/droid_oxe/Put_the_yellow_object_yellow_pencil_and_pink_object_inside_the_open_cabinet/episode_0/cam_right_wrist/frame_000435.jpg", "images/droid_oxe/Put_the_yellow_object_yellow_pencil_and_pink_object_inside_the_open_cabinet/episode_0/cam_high/frame_000360.jpg", "images/droid_oxe/Put_the_yellow_object_yellow_pencil_and_pink_object_inside_the_open_cabinet/episode_0/cam_left_wrist/frame_000360.jpg", "images/droid_oxe/Put_the_yellow_object_yellow_pencil_and_pink_object_inside_the_open_cabinet/episode_0/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the yellow object yellow pencil and pink object inside the open cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the yellow object yellow pencil and pink object inside the open cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -17.2, "hop_absolute_value": 17.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Empty_the_contents_on_the_plate_into_the_bin_episode_0-train-bf_000540-af_000315-minus-0000225", "task": "Empty the contents on the plate into the bin", "image": [ "images/droid_oxe/Empty_the_contents_on_the_plate_into_the_bin/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Empty_the_contents_on_the_plate_into_the_bin/episode_0/cam_high/frame_000580.jpg", "images/droid_oxe/Empty_the_contents_on_the_plate_into_the_bin/episode_0/cam_high/frame_000540.jpg", "images/droid_oxe/Empty_the_contents_on_the_plate_into_the_bin/episode_0/cam_left_wrist/frame_000540.jpg", "images/droid_oxe/Empty_the_contents_on_the_plate_into_the_bin/episode_0/cam_right_wrist/frame_000540.jpg", "images/droid_oxe/Empty_the_contents_on_the_plate_into_the_bin/episode_0/cam_high/frame_000315.jpg", "images/droid_oxe/Empty_the_contents_on_the_plate_into_the_bin/episode_0/cam_left_wrist/frame_000315.jpg", "images/droid_oxe/Empty_the_contents_on_the_plate_into_the_bin/episode_0/cam_right_wrist/frame_000315.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Empty the contents on the plate into the bin`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Empty the contents on the plate into the bin`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -41.7, "hop_absolute_value": 41.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pour_the_contents_in_the_clear_jug_into_the_red_bowl_episode_6-train-bf_000240-af_000015-minus-0000020", "task": "Pour the contents in the clear jug into the red bowl", "image": [ "images/droid_oxe/Pour_the_contents_in_the_clear_jug_into_the_red_bowl/episode_6/cam_high/frame_000000.jpg", "images/droid_oxe/Pour_the_contents_in_the_clear_jug_into_the_red_bowl/episode_6/cam_high/frame_000364.jpg", "images/droid_oxe/Pour_the_contents_in_the_clear_jug_into_the_red_bowl/episode_6/cam_high/frame_000240.jpg", "images/droid_oxe/Pour_the_contents_in_the_clear_jug_into_the_red_bowl/episode_6/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Pour_the_contents_in_the_clear_jug_into_the_red_bowl/episode_6/cam_right_wrist/frame_000240.jpg", "images/droid_oxe/Pour_the_contents_in_the_clear_jug_into_the_red_bowl/episode_6/cam_high/frame_000015.jpg", "images/droid_oxe/Pour_the_contents_in_the_clear_jug_into_the_red_bowl/episode_6/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Pour_the_contents_in_the_clear_jug_into_the_red_bowl/episode_6/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pour the contents in the clear jug into the red bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pour the contents in the clear jug into the red bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -93.8, "hop_absolute_value": 93.8, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Open_the_laptop_episode_2-train-bf_000150-af_000045-minus-0000033", "task": "Open the laptop", "image": [ "images/droid_oxe/Open_the_laptop/episode_11/cam_high/frame_000000.jpg", "images/droid_oxe/Open_the_laptop/episode_11/cam_high/frame_000151.jpg", "images/droid_oxe/Open_the_laptop/episode_2/cam_high/frame_000150.jpg", "images/droid_oxe/Open_the_laptop/episode_2/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Open_the_laptop/episode_2/cam_right_wrist/frame_000150.jpg", "images/droid_oxe/Open_the_laptop/episode_2/cam_high/frame_000045.jpg", "images/droid_oxe/Open_the_laptop/episode_2/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Open_the_laptop/episode_2/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the laptop`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the laptop`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_cloths_and_put_them_in_the_bag_episode_0-train-bf_000120-af_000030-minus-0000039", "task": "Pick up the cloths and put them in the bag", "image": [ "images/droid_oxe/Pick_up_the_cloths_and_put_them_in_the_bag/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_cloths_and_put_them_in_the_bag/episode_0/cam_high/frame_000231.jpg", "images/droid_oxe/Pick_up_the_cloths_and_put_them_in_the_bag/episode_0/cam_high/frame_000120.jpg", "images/droid_oxe/Pick_up_the_cloths_and_put_them_in_the_bag/episode_0/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Pick_up_the_cloths_and_put_them_in_the_bag/episode_0/cam_right_wrist/frame_000120.jpg", "images/droid_oxe/Pick_up_the_cloths_and_put_them_in_the_bag/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Pick_up_the_cloths_and_put_them_in_the_bag/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Pick_up_the_cloths_and_put_them_in_the_bag/episode_0/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the cloths and put them in the bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the cloths and put them in the bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_purple_plush_toy_on_the_beige_plate_episode_2-train-bf_000105-af_000000-minus-0000006", "task": "Put the purple plush toy on the beige plate", "image": [ "images/droid_oxe/Put_the_purple_plush_toy_on_the_beige_plate/episode_3/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_purple_plush_toy_on_the_beige_plate/episode_3/cam_high/frame_000126.jpg", "images/droid_oxe/Put_the_purple_plush_toy_on_the_beige_plate/episode_2/cam_high/frame_000105.jpg", "images/droid_oxe/Put_the_purple_plush_toy_on_the_beige_plate/episode_2/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Put_the_purple_plush_toy_on_the_beige_plate/episode_2/cam_right_wrist/frame_000105.jpg", "images/droid_oxe/Put_the_purple_plush_toy_on_the_beige_plate/episode_2/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_purple_plush_toy_on_the_beige_plate/episode_2/cam_left_wrist/frame_000000.jpg", "images/droid_oxe/Put_the_purple_plush_toy_on_the_beige_plate/episode_2/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the purple plush toy on the beige plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the purple plush toy on the beige plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_spatula_to_mix_the_contents_of_the_pot_then_put_it_back_on_the_table_episode_0-train-bf_000360-af_000210-minus-0000219", "task": "Use the spatula to mix the contents of the pot then put it back on the table", "image": [ "images/droid_oxe/Use_the_spatula_to_mix_the_contents_of_the_pot_then_put_it_back_on_the_table/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_spatula_to_mix_the_contents_of_the_pot_then_put_it_back_on_the_table/episode_0/cam_high/frame_000497.jpg", "images/droid_oxe/Use_the_spatula_to_mix_the_contents_of_the_pot_then_put_it_back_on_the_table/episode_0/cam_high/frame_000360.jpg", "images/droid_oxe/Use_the_spatula_to_mix_the_contents_of_the_pot_then_put_it_back_on_the_table/episode_0/cam_left_wrist/frame_000360.jpg", "images/droid_oxe/Use_the_spatula_to_mix_the_contents_of_the_pot_then_put_it_back_on_the_table/episode_0/cam_right_wrist/frame_000360.jpg", "images/droid_oxe/Use_the_spatula_to_mix_the_contents_of_the_pot_then_put_it_back_on_the_table/episode_0/cam_high/frame_000210.jpg", "images/droid_oxe/Use_the_spatula_to_mix_the_contents_of_the_pot_then_put_it_back_on_the_table/episode_0/cam_left_wrist/frame_000210.jpg", "images/droid_oxe/Use_the_spatula_to_mix_the_contents_of_the_pot_then_put_it_back_on_the_table/episode_0/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the spatula to mix the contents of the pot then put it back on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the spatula to mix the contents of the pot then put it back on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -41.7, "hop_absolute_value": 41.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Open_the_upper_part_of_the_cabinet_door_episode_0-train-bf_000285-af_000030-minus-0000049", "task": "Open the upper part of the cabinet door", "image": [ "images/droid_oxe/Open_the_upper_part_of_the_cabinet_door/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Open_the_upper_part_of_the_cabinet_door/episode_0/cam_high/frame_000652.jpg", "images/droid_oxe/Open_the_upper_part_of_the_cabinet_door/episode_0/cam_high/frame_000285.jpg", "images/droid_oxe/Open_the_upper_part_of_the_cabinet_door/episode_0/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/Open_the_upper_part_of_the_cabinet_door/episode_0/cam_right_wrist/frame_000285.jpg", "images/droid_oxe/Open_the_upper_part_of_the_cabinet_door/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Open_the_upper_part_of_the_cabinet_door/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Open_the_upper_part_of_the_cabinet_door/episode_0/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the upper part of the cabinet door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the upper part of the cabinet door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -89.5, "hop_absolute_value": 89.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_yellow_vessel_from_the_cabinet_and_put_it_on_the_counter_episode_0-train-bf_000315-af_000180-minus-0000679", "task": "Remove the yellow vessel from the cabinet and put it on the counter", "image": [ "images/droid_oxe/Remove_the_yellow_vessel_from_the_cabinet_and_put_it_on_the_counter/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_yellow_vessel_from_the_cabinet_and_put_it_on_the_counter/episode_0/cam_high/frame_000432.jpg", "images/droid_oxe/Remove_the_yellow_vessel_from_the_cabinet_and_put_it_on_the_counter/episode_0/cam_high/frame_000315.jpg", "images/droid_oxe/Remove_the_yellow_vessel_from_the_cabinet_and_put_it_on_the_counter/episode_0/cam_left_wrist/frame_000315.jpg", "images/droid_oxe/Remove_the_yellow_vessel_from_the_cabinet_and_put_it_on_the_counter/episode_0/cam_right_wrist/frame_000315.jpg", "images/droid_oxe/Remove_the_yellow_vessel_from_the_cabinet_and_put_it_on_the_counter/episode_0/cam_high/frame_000180.jpg", "images/droid_oxe/Remove_the_yellow_vessel_from_the_cabinet_and_put_it_on_the_counter/episode_0/cam_left_wrist/frame_000180.jpg", "images/droid_oxe/Remove_the_yellow_vessel_from_the_cabinet_and_put_it_on_the_counter/episode_0/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the yellow vessel from the cabinet and put it on the counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the yellow vessel from the cabinet and put it on the counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -42.9, "hop_absolute_value": 42.9, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_round_blue_block_on_the_table_episode_0-train-bf_000270-af_000120-minus-0000458", "task": "Put the round blue block on the table", "image": [ "images/droid_oxe/Put_the_round_blue_block_on_the_table/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_round_blue_block_on_the_table/episode_0/cam_high/frame_000316.jpg", "images/droid_oxe/Put_the_round_blue_block_on_the_table/episode_0/cam_high/frame_000270.jpg", "images/droid_oxe/Put_the_round_blue_block_on_the_table/episode_0/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Put_the_round_blue_block_on_the_table/episode_0/cam_right_wrist/frame_000270.jpg", "images/droid_oxe/Put_the_round_blue_block_on_the_table/episode_0/cam_high/frame_000120.jpg", "images/droid_oxe/Put_the_round_blue_block_on_the_table/episode_0/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Put_the_round_blue_block_on_the_table/episode_0/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the round blue block on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the round blue block on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -55.6, "hop_absolute_value": 55.6, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_blue_shirt_in_the_blue_laundry_bag_episode_0-train-bf_000225-af_000075-minus-0000128", "task": "Put the blue shirt in the blue laundry bag", "image": [ "images/droid_oxe/Put_the_blue_shirt_in_the_blue_laundry_bag/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_blue_shirt_in_the_blue_laundry_bag/episode_0/cam_high/frame_000420.jpg", "images/droid_oxe/Put_the_blue_shirt_in_the_blue_laundry_bag/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Put_the_blue_shirt_in_the_blue_laundry_bag/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_blue_shirt_in_the_blue_laundry_bag/episode_0/cam_right_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_blue_shirt_in_the_blue_laundry_bag/episode_0/cam_high/frame_000075.jpg", "images/droid_oxe/Put_the_blue_shirt_in_the_blue_laundry_bag/episode_0/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Put_the_blue_shirt_in_the_blue_laundry_bag/episode_0/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the blue shirt in the blue laundry bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the blue shirt in the blue laundry bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_lid_and_close_the_bottle_episode_0-train-bf_000105-af_000015-minus-0000027", "task": "Pick up the lid and close the bottle", "image": [ "images/droid_oxe/Pick_up_the_lid_and_close_the_bottle/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_lid_and_close_the_bottle/episode_1/cam_high/frame_000203.jpg", "images/droid_oxe/Pick_up_the_lid_and_close_the_bottle/episode_0/cam_high/frame_000105.jpg", "images/droid_oxe/Pick_up_the_lid_and_close_the_bottle/episode_0/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Pick_up_the_lid_and_close_the_bottle/episode_0/cam_right_wrist/frame_000105.jpg", "images/droid_oxe/Pick_up_the_lid_and_close_the_bottle/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Pick_up_the_lid_and_close_the_bottle/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Pick_up_the_lid_and_close_the_bottle/episode_0/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the lid and close the bottle`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the lid and close the bottle`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_bottle_with_the_yellow_lid_to_the_upper_right_corner_of_the_counter_episode_0-train-bf_000225-af_000210-minus-0000183", "task": "Move the bottle with the yellow lid to the upper right corner of the counter", "image": [ "images/droid_oxe/Move_the_bottle_with_the_yellow_lid_to_the_upper_right_corner_of_the_counter/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_bottle_with_the_yellow_lid_to_the_upper_right_corner_of_the_counter/episode_0/cam_high/frame_000308.jpg", "images/droid_oxe/Move_the_bottle_with_the_yellow_lid_to_the_upper_right_corner_of_the_counter/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Move_the_bottle_with_the_yellow_lid_to_the_upper_right_corner_of_the_counter/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Move_the_bottle_with_the_yellow_lid_to_the_upper_right_corner_of_the_counter/episode_0/cam_right_wrist/frame_000225.jpg", "images/droid_oxe/Move_the_bottle_with_the_yellow_lid_to_the_upper_right_corner_of_the_counter/episode_0/cam_high/frame_000210.jpg", "images/droid_oxe/Move_the_bottle_with_the_yellow_lid_to_the_upper_right_corner_of_the_counter/episode_0/cam_left_wrist/frame_000210.jpg", "images/droid_oxe/Move_the_bottle_with_the_yellow_lid_to_the_upper_right_corner_of_the_counter/episode_0/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the bottle with the yellow lid to the upper right corner of the counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the bottle with the yellow lid to the upper right corner of the counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.7, "hop_absolute_value": 6.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Place_the_pink_clothing_inside_the_black_bag_episode_0-train-bf_000270-af_000255-minus-0000403", "task": "Place the pink clothing inside the black bag", "image": [ "images/droid_oxe/Place_the_pink_clothing_inside_the_black_bag/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Place_the_pink_clothing_inside_the_black_bag/episode_0/cam_high/frame_000327.jpg", "images/droid_oxe/Place_the_pink_clothing_inside_the_black_bag/episode_0/cam_high/frame_000270.jpg", "images/droid_oxe/Place_the_pink_clothing_inside_the_black_bag/episode_0/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Place_the_pink_clothing_inside_the_black_bag/episode_0/cam_right_wrist/frame_000270.jpg", "images/droid_oxe/Place_the_pink_clothing_inside_the_black_bag/episode_0/cam_high/frame_000255.jpg", "images/droid_oxe/Place_the_pink_clothing_inside_the_black_bag/episode_0/cam_left_wrist/frame_000255.jpg", "images/droid_oxe/Place_the_pink_clothing_inside_the_black_bag/episode_0/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the pink clothing inside the black bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the pink clothing inside the black bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.6, "hop_absolute_value": 5.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_plush_toy_episode_5-train-bf_000330-af_000090-minus-0000120", "task": "Move the plush toy", "image": [ "images/droid_oxe/Move_the_plush_toy/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_plush_toy/episode_0/cam_high/frame_000116.jpg", "images/droid_oxe/Move_the_plush_toy/episode_5/cam_high/frame_000330.jpg", "images/droid_oxe/Move_the_plush_toy/episode_5/cam_left_wrist/frame_000330.jpg", "images/droid_oxe/Move_the_plush_toy/episode_5/cam_right_wrist/frame_000330.jpg", "images/droid_oxe/Move_the_plush_toy/episode_5/cam_high/frame_000090.jpg", "images/droid_oxe/Move_the_plush_toy/episode_5/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Move_the_plush_toy/episode_5/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the plush toy`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the plush toy`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -72.7, "hop_absolute_value": 72.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Open_the_cabinet_episode_10-train-bf_000210-af_000195-minus-0000238", "task": "Open the cabinet", "image": [ "images/droid_oxe/Open_the_cabinet/episode_14/cam_high/frame_000000.jpg", "images/droid_oxe/Open_the_cabinet/episode_14/cam_high/frame_000409.jpg", "images/droid_oxe/Open_the_cabinet/episode_10/cam_high/frame_000210.jpg", "images/droid_oxe/Open_the_cabinet/episode_10/cam_left_wrist/frame_000210.jpg", "images/droid_oxe/Open_the_cabinet/episode_10/cam_right_wrist/frame_000210.jpg", "images/droid_oxe/Open_the_cabinet/episode_10/cam_high/frame_000195.jpg", "images/droid_oxe/Open_the_cabinet/episode_10/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Open_the_cabinet/episode_10/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.1, "hop_absolute_value": 7.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Take_the_fabric_softener_and_pour_it_in_the_washing_machine_episode_0-train-bf_000225-af_000045-minus-0000059", "task": "Take the fabric softener and pour it in the washing machine", "image": [ "images/droid_oxe/Take_the_fabric_softener_and_pour_it_in_the_washing_machine/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Take_the_fabric_softener_and_pour_it_in_the_washing_machine/episode_0/cam_high/frame_000324.jpg", "images/droid_oxe/Take_the_fabric_softener_and_pour_it_in_the_washing_machine/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Take_the_fabric_softener_and_pour_it_in_the_washing_machine/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Take_the_fabric_softener_and_pour_it_in_the_washing_machine/episode_0/cam_right_wrist/frame_000225.jpg", "images/droid_oxe/Take_the_fabric_softener_and_pour_it_in_the_washing_machine/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Take_the_fabric_softener_and_pour_it_in_the_washing_machine/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Take_the_fabric_softener_and_pour_it_in_the_washing_machine/episode_0/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the fabric softener and pour it in the washing machine`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the fabric softener and pour it in the washing machine`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_the_cloth_and_wipe_the_sauce_pan_episode_0-train-bf_000210-af_000030-minus-0000043", "task": "Pick the cloth and wipe the sauce pan", "image": [ "images/droid_oxe/Pick_the_cloth_and_wipe_the_sauce_pan/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_the_cloth_and_wipe_the_sauce_pan/episode_0/cam_high/frame_000350.jpg", "images/droid_oxe/Pick_the_cloth_and_wipe_the_sauce_pan/episode_0/cam_high/frame_000210.jpg", "images/droid_oxe/Pick_the_cloth_and_wipe_the_sauce_pan/episode_0/cam_left_wrist/frame_000210.jpg", "images/droid_oxe/Pick_the_cloth_and_wipe_the_sauce_pan/episode_0/cam_right_wrist/frame_000210.jpg", "images/droid_oxe/Pick_the_cloth_and_wipe_the_sauce_pan/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Pick_the_cloth_and_wipe_the_sauce_pan/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Pick_the_cloth_and_wipe_the_sauce_pan/episode_0/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the cloth and wipe the sauce pan`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the cloth and wipe the sauce pan`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_bottle_from_the_bowl_to_the_table_episode_0-train-bf_000195-af_000180-minus-0000264", "task": "Move the bottle from the bowl to the table", "image": [ "images/droid_oxe/Move_the_bottle_from_the_bowl_to_the_table/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_bottle_from_the_bowl_to_the_table/episode_0/cam_high/frame_000474.jpg", "images/droid_oxe/Move_the_bottle_from_the_bowl_to_the_table/episode_0/cam_high/frame_000195.jpg", "images/droid_oxe/Move_the_bottle_from_the_bowl_to_the_table/episode_0/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Move_the_bottle_from_the_bowl_to_the_table/episode_0/cam_right_wrist/frame_000195.jpg", "images/droid_oxe/Move_the_bottle_from_the_bowl_to_the_table/episode_0/cam_high/frame_000180.jpg", "images/droid_oxe/Move_the_bottle_from_the_bowl_to_the_table/episode_0/cam_left_wrist/frame_000180.jpg", "images/droid_oxe/Move_the_bottle_from_the_bowl_to_the_table/episode_0/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the bottle from the bowl to the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the bottle from the bowl to the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.7, "hop_absolute_value": 7.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_a_book_from_the_second_shelf_from_the_right_episode_0-train-bf_000135-af_000045-minus-0000109", "task": "Remove a book from the second shelf from the right", "image": [ "images/droid_oxe/Remove_a_book_from_the_second_shelf_from_the_right/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_a_book_from_the_second_shelf_from_the_right/episode_0/cam_high/frame_000375.jpg", "images/droid_oxe/Remove_a_book_from_the_second_shelf_from_the_right/episode_0/cam_high/frame_000135.jpg", "images/droid_oxe/Remove_a_book_from_the_second_shelf_from_the_right/episode_0/cam_left_wrist/frame_000135.jpg", "images/droid_oxe/Remove_a_book_from_the_second_shelf_from_the_right/episode_0/cam_right_wrist/frame_000135.jpg", "images/droid_oxe/Remove_a_book_from_the_second_shelf_from_the_right/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Remove_a_book_from_the_second_shelf_from_the_right/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Remove_a_book_from_the_second_shelf_from_the_right/episode_0/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove a book from the second shelf from the right`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove a book from the second shelf from the right`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_the_kitchen_towel_and_wipe_the_pan_episode_0-train-bf_001665-af_000390-minus-0000097", "task": "Pick up the kitchen towel and wipe the pan", "image": [ "images/droid_oxe/Pick_up_the_kitchen_towel_and_wipe_the_pan/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_the_kitchen_towel_and_wipe_the_pan/episode_0/cam_high/frame_002061.jpg", "images/droid_oxe/Pick_up_the_kitchen_towel_and_wipe_the_pan/episode_0/cam_high/frame_001665.jpg", "images/droid_oxe/Pick_up_the_kitchen_towel_and_wipe_the_pan/episode_0/cam_left_wrist/frame_001665.jpg", "images/droid_oxe/Pick_up_the_kitchen_towel_and_wipe_the_pan/episode_0/cam_right_wrist/frame_001665.jpg", "images/droid_oxe/Pick_up_the_kitchen_towel_and_wipe_the_pan/episode_0/cam_high/frame_000390.jpg", "images/droid_oxe/Pick_up_the_kitchen_towel_and_wipe_the_pan/episode_0/cam_left_wrist/frame_000390.jpg", "images/droid_oxe/Pick_up_the_kitchen_towel_and_wipe_the_pan/episode_0/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the kitchen towel and wipe the pan`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the kitchen towel and wipe the pan`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -76.6, "hop_absolute_value": 76.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_flask_inside_the_flask_and_use_it_to_scrub_the_inside_episode_0-train-bf_000240-af_000195-minus-0000556", "task": "Put the flask inside the flask and use it to scrub the inside", "image": [ "images/droid_oxe/Put_the_flask_inside_the_flask_and_use_it_to_scrub_the_inside/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_flask_inside_the_flask_and_use_it_to_scrub_the_inside/episode_0/cam_high/frame_000412.jpg", "images/droid_oxe/Put_the_flask_inside_the_flask_and_use_it_to_scrub_the_inside/episode_0/cam_high/frame_000240.jpg", "images/droid_oxe/Put_the_flask_inside_the_flask_and_use_it_to_scrub_the_inside/episode_0/cam_left_wrist/frame_000240.jpg", "images/droid_oxe/Put_the_flask_inside_the_flask_and_use_it_to_scrub_the_inside/episode_0/cam_right_wrist/frame_000240.jpg", "images/droid_oxe/Put_the_flask_inside_the_flask_and_use_it_to_scrub_the_inside/episode_0/cam_high/frame_000195.jpg", "images/droid_oxe/Put_the_flask_inside_the_flask_and_use_it_to_scrub_the_inside/episode_0/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Put_the_flask_inside_the_flask_and_use_it_to_scrub_the_inside/episode_0/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the flask inside the flask and use it to scrub the inside`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the flask inside the flask and use it to scrub the inside`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -18.8, "hop_absolute_value": 18.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Open_the_white_and_green_bin_episode_0-train-bf_000435-af_000420-minus-0000589", "task": "Open the white and green bin", "image": [ "images/droid_oxe/Open_the_white_and_green_bin/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Open_the_white_and_green_bin/episode_0/cam_high/frame_000569.jpg", "images/droid_oxe/Open_the_white_and_green_bin/episode_0/cam_high/frame_000435.jpg", "images/droid_oxe/Open_the_white_and_green_bin/episode_0/cam_left_wrist/frame_000435.jpg", "images/droid_oxe/Open_the_white_and_green_bin/episode_0/cam_right_wrist/frame_000435.jpg", "images/droid_oxe/Open_the_white_and_green_bin/episode_0/cam_high/frame_000420.jpg", "images/droid_oxe/Open_the_white_and_green_bin/episode_0/cam_left_wrist/frame_000420.jpg", "images/droid_oxe/Open_the_white_and_green_bin/episode_0/cam_right_wrist/frame_000420.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the white and green bin`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the white and green bin`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.4, "hop_absolute_value": 3.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Close_the_bottom_drawer_of_the_file_cabinet_episode_14-train-bf_000045-af_000030-minus-0000008", "task": "Close the bottom drawer of the file cabinet", "image": [ "images/droid_oxe/Close_the_bottom_drawer_of_the_file_cabinet/episode_9/cam_high/frame_000000.jpg", "images/droid_oxe/Close_the_bottom_drawer_of_the_file_cabinet/episode_9/cam_high/frame_000070.jpg", "images/droid_oxe/Close_the_bottom_drawer_of_the_file_cabinet/episode_14/cam_high/frame_000045.jpg", "images/droid_oxe/Close_the_bottom_drawer_of_the_file_cabinet/episode_14/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Close_the_bottom_drawer_of_the_file_cabinet/episode_14/cam_right_wrist/frame_000045.jpg", "images/droid_oxe/Close_the_bottom_drawer_of_the_file_cabinet/episode_14/cam_high/frame_000030.jpg", "images/droid_oxe/Close_the_bottom_drawer_of_the_file_cabinet/episode_14/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Close_the_bottom_drawer_of_the_file_cabinet/episode_14/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the bottom drawer of the file cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the bottom drawer of the file cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Use_the_towel_to_wipe_the_whiteboard_episode_0-train-bf_000195-af_000045-minus-0000075", "task": "Use the towel to wipe the whiteboard", "image": [ "images/droid_oxe/Use_the_towel_to_wipe_the_whiteboard/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Use_the_towel_to_wipe_the_whiteboard/episode_1/cam_high/frame_000758.jpg", "images/droid_oxe/Use_the_towel_to_wipe_the_whiteboard/episode_0/cam_high/frame_000195.jpg", "images/droid_oxe/Use_the_towel_to_wipe_the_whiteboard/episode_0/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Use_the_towel_to_wipe_the_whiteboard/episode_0/cam_right_wrist/frame_000195.jpg", "images/droid_oxe/Use_the_towel_to_wipe_the_whiteboard/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Use_the_towel_to_wipe_the_whiteboard/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Use_the_towel_to_wipe_the_whiteboard/episode_0/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the towel to wipe the whiteboard`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the towel to wipe the whiteboard`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -76.9, "hop_absolute_value": 76.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_lid_from_the_jug_episode_0-train-bf_000165-af_000150-minus-0000165", "task": "Remove the lid from the jug", "image": [ "images/droid_oxe/Remove_the_lid_from_the_jug/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_lid_from_the_jug/episode_0/cam_high/frame_000291.jpg", "images/droid_oxe/Remove_the_lid_from_the_jug/episode_0/cam_high/frame_000165.jpg", "images/droid_oxe/Remove_the_lid_from_the_jug/episode_0/cam_left_wrist/frame_000165.jpg", "images/droid_oxe/Remove_the_lid_from_the_jug/episode_0/cam_right_wrist/frame_000165.jpg", "images/droid_oxe/Remove_the_lid_from_the_jug/episode_0/cam_high/frame_000150.jpg", "images/droid_oxe/Remove_the_lid_from_the_jug/episode_0/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Remove_the_lid_from_the_jug/episode_0/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the lid from the jug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the lid from the jug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Push_the_smaller_bottle_down_episode_0-train-bf_000030-af_000015-minus-0000031", "task": "Push the smaller bottle down", "image": [ "images/droid_oxe/Push_the_smaller_bottle_down/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Push_the_smaller_bottle_down/episode_0/cam_high/frame_000157.jpg", "images/droid_oxe/Push_the_smaller_bottle_down/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Push_the_smaller_bottle_down/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Push_the_smaller_bottle_down/episode_0/cam_right_wrist/frame_000030.jpg", "images/droid_oxe/Push_the_smaller_bottle_down/episode_0/cam_high/frame_000015.jpg", "images/droid_oxe/Push_the_smaller_bottle_down/episode_0/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Push_the_smaller_bottle_down/episode_0/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the smaller bottle down`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the smaller bottle down`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_fork_to_the_right_episode_1-train-bf_000255-af_000045-minus-0000057", "task": "Move the fork to the right", "image": [ "images/droid_oxe/Move_the_fork_to_the_right/episode_1/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_fork_to_the_right/episode_1/cam_high/frame_000328.jpg", "images/droid_oxe/Move_the_fork_to_the_right/episode_1/cam_high/frame_000255.jpg", "images/droid_oxe/Move_the_fork_to_the_right/episode_1/cam_left_wrist/frame_000255.jpg", "images/droid_oxe/Move_the_fork_to_the_right/episode_1/cam_right_wrist/frame_000255.jpg", "images/droid_oxe/Move_the_fork_to_the_right/episode_1/cam_high/frame_000045.jpg", "images/droid_oxe/Move_the_fork_to_the_right/episode_1/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Move_the_fork_to_the_right/episode_1/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the fork to the right`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the fork to the right`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -82.4, "hop_absolute_value": 82.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Take_the_gluestick_out_of_the_green_cup_and_place_it_on_the_silver_object_episode_0-train-bf_000210-af_000195-minus-0000144", "task": "Take the gluestick out of the green cup and place it on the silver object", "image": [ "images/droid_oxe/Take_the_gluestick_out_of_the_green_cup_and_place_it_on_the_silver_object/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Take_the_gluestick_out_of_the_green_cup_and_place_it_on_the_silver_object/episode_0/cam_high/frame_000263.jpg", "images/droid_oxe/Take_the_gluestick_out_of_the_green_cup_and_place_it_on_the_silver_object/episode_0/cam_high/frame_000210.jpg", "images/droid_oxe/Take_the_gluestick_out_of_the_green_cup_and_place_it_on_the_silver_object/episode_0/cam_left_wrist/frame_000210.jpg", "images/droid_oxe/Take_the_gluestick_out_of_the_green_cup_and_place_it_on_the_silver_object/episode_0/cam_right_wrist/frame_000210.jpg", "images/droid_oxe/Take_the_gluestick_out_of_the_green_cup_and_place_it_on_the_silver_object/episode_0/cam_high/frame_000195.jpg", "images/droid_oxe/Take_the_gluestick_out_of_the_green_cup_and_place_it_on_the_silver_object/episode_0/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Take_the_gluestick_out_of_the_green_cup_and_place_it_on_the_silver_object/episode_0/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the gluestick out of the green cup and place it on the silver object`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the gluestick out of the green cup and place it on the silver object`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.1, "hop_absolute_value": 7.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_remote_control_on_top_of_the_shelf_episode_0-train-bf_000225-af_000120-minus-0000103", "task": "Put the remote control on top of the shelf", "image": [ "images/droid_oxe/Put_the_remote_control_on_top_of_the_shelf/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_remote_control_on_top_of_the_shelf/episode_0/cam_high/frame_000277.jpg", "images/droid_oxe/Put_the_remote_control_on_top_of_the_shelf/episode_0/cam_high/frame_000225.jpg", "images/droid_oxe/Put_the_remote_control_on_top_of_the_shelf/episode_0/cam_left_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_remote_control_on_top_of_the_shelf/episode_0/cam_right_wrist/frame_000225.jpg", "images/droid_oxe/Put_the_remote_control_on_top_of_the_shelf/episode_0/cam_high/frame_000120.jpg", "images/droid_oxe/Put_the_remote_control_on_top_of_the_shelf/episode_0/cam_left_wrist/frame_000120.jpg", "images/droid_oxe/Put_the_remote_control_on_top_of_the_shelf/episode_0/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the remote control on top of the shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the remote control on top of the shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -46.7, "hop_absolute_value": 46.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_a_can_from_the_middle_to_the_bottom_shelf_episode_0-train-bf_000285-af_000045-minus-0000051", "task": "Remove a can from the middle to the bottom shelf", "image": [ "images/droid_oxe/Remove_a_can_from_the_middle_to_the_bottom_shelf/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_a_can_from_the_middle_to_the_bottom_shelf/episode_0/cam_high/frame_000402.jpg", "images/droid_oxe/Remove_a_can_from_the_middle_to_the_bottom_shelf/episode_0/cam_high/frame_000285.jpg", "images/droid_oxe/Remove_a_can_from_the_middle_to_the_bottom_shelf/episode_0/cam_left_wrist/frame_000285.jpg", "images/droid_oxe/Remove_a_can_from_the_middle_to_the_bottom_shelf/episode_0/cam_right_wrist/frame_000285.jpg", "images/droid_oxe/Remove_a_can_from_the_middle_to_the_bottom_shelf/episode_0/cam_high/frame_000045.jpg", "images/droid_oxe/Remove_a_can_from_the_middle_to_the_bottom_shelf/episode_0/cam_left_wrist/frame_000045.jpg", "images/droid_oxe/Remove_a_can_from_the_middle_to_the_bottom_shelf/episode_0/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove a can from the middle to the bottom shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove a can from the middle to the bottom shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -84.2, "hop_absolute_value": 84.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Press_on_the_keyboard_episode_2-train-bf_000250-af_000150-minus-0000091", "task": "Press on the keyboard", "image": [ "images/droid_oxe/Press_on_the_keyboard/episode_2/cam_high/frame_000000.jpg", "images/droid_oxe/Press_on_the_keyboard/episode_2/cam_high/frame_000250.jpg", "images/droid_oxe/Press_on_the_keyboard/episode_2/cam_high/frame_000250.jpg", "images/droid_oxe/Press_on_the_keyboard/episode_2/cam_left_wrist/frame_000250.jpg", "images/droid_oxe/Press_on_the_keyboard/episode_2/cam_right_wrist/frame_000250.jpg", "images/droid_oxe/Press_on_the_keyboard/episode_2/cam_high/frame_000150.jpg", "images/droid_oxe/Press_on_the_keyboard/episode_2/cam_left_wrist/frame_000150.jpg", "images/droid_oxe/Press_on_the_keyboard/episode_2/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Press on the keyboard`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Press on the keyboard`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -37.5, "hop_absolute_value": 37.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Move_the_mouse_to_the_bottom_of_the_table_the_take_the_two_cans_and_put_them_on_the_paper_episode_0-train-bf_000090-af_000030-minus-0000136", "task": "Move the mouse to the bottom of the table the take the two cans and put them on the paper", "image": [ "images/droid_oxe/Move_the_mouse_to_the_bottom_of_the_table_the_take_the_two_cans_and_put_them_on_the_paper/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Move_the_mouse_to_the_bottom_of_the_table_the_take_the_two_cans_and_put_them_on_the_paper/episode_0/cam_high/frame_000595.jpg", "images/droid_oxe/Move_the_mouse_to_the_bottom_of_the_table_the_take_the_two_cans_and_put_them_on_the_paper/episode_0/cam_high/frame_000090.jpg", "images/droid_oxe/Move_the_mouse_to_the_bottom_of_the_table_the_take_the_two_cans_and_put_them_on_the_paper/episode_0/cam_left_wrist/frame_000090.jpg", "images/droid_oxe/Move_the_mouse_to_the_bottom_of_the_table_the_take_the_two_cans_and_put_them_on_the_paper/episode_0/cam_right_wrist/frame_000090.jpg", "images/droid_oxe/Move_the_mouse_to_the_bottom_of_the_table_the_take_the_two_cans_and_put_them_on_the_paper/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Move_the_mouse_to_the_bottom_of_the_table_the_take_the_two_cans_and_put_them_on_the_paper/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Move_the_mouse_to_the_bottom_of_the_table_the_take_the_two_cans_and_put_them_on_the_paper/episode_0/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the mouse to the bottom of the table the take the two cans and put them on the paper`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the mouse to the bottom of the table the take the two cans and put them on the paper`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Pick_up_two_small_books_from_the_compartment_of_the_book_shelf_and_move_them_to_the_right_episode_0-train-bf_000645-af_000330-minus-0000220", "task": "Pick up two small books from the compartment of the book shelf and move them to the right", "image": [ "images/droid_oxe/Pick_up_two_small_books_from_the_compartment_of_the_book_shelf_and_move_them_to_the_right/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Pick_up_two_small_books_from_the_compartment_of_the_book_shelf_and_move_them_to_the_right/episode_0/cam_high/frame_000955.jpg", "images/droid_oxe/Pick_up_two_small_books_from_the_compartment_of_the_book_shelf_and_move_them_to_the_right/episode_0/cam_high/frame_000645.jpg", "images/droid_oxe/Pick_up_two_small_books_from_the_compartment_of_the_book_shelf_and_move_them_to_the_right/episode_0/cam_left_wrist/frame_000645.jpg", "images/droid_oxe/Pick_up_two_small_books_from_the_compartment_of_the_book_shelf_and_move_them_to_the_right/episode_0/cam_right_wrist/frame_000645.jpg", "images/droid_oxe/Pick_up_two_small_books_from_the_compartment_of_the_book_shelf_and_move_them_to_the_right/episode_0/cam_high/frame_000330.jpg", "images/droid_oxe/Pick_up_two_small_books_from_the_compartment_of_the_book_shelf_and_move_them_to_the_right/episode_0/cam_left_wrist/frame_000330.jpg", "images/droid_oxe/Pick_up_two_small_books_from_the_compartment_of_the_book_shelf_and_move_them_to_the_right/episode_0/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up two small books from the compartment of the book shelf and move them to the right`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up two small books from the compartment of the book shelf and move them to the right`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -48.8, "hop_absolute_value": 48.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Press_a_button_on_the_laptop_episode_1-train-bf_000096-af_000015-minus-0000007", "task": "Press a button on the laptop", "image": [ "images/droid_oxe/Press_a_button_on_the_laptop/episode_2/cam_high/frame_000000.jpg", "images/droid_oxe/Press_a_button_on_the_laptop/episode_2/cam_high/frame_000154.jpg", "images/droid_oxe/Press_a_button_on_the_laptop/episode_1/cam_high/frame_000096.jpg", "images/droid_oxe/Press_a_button_on_the_laptop/episode_1/cam_left_wrist/frame_000096.jpg", "images/droid_oxe/Press_a_button_on_the_laptop/episode_1/cam_right_wrist/frame_000096.jpg", "images/droid_oxe/Press_a_button_on_the_laptop/episode_1/cam_high/frame_000015.jpg", "images/droid_oxe/Press_a_button_on_the_laptop/episode_1/cam_left_wrist/frame_000015.jpg", "images/droid_oxe/Press_a_button_on_the_laptop/episode_1/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Press a button on the laptop`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Press a button on the laptop`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Switch_off_the_extension_cord_episode_8-train-bf_000180-af_000075-minus-0000060", "task": "Switch off the extension cord", "image": [ "images/droid_oxe/Switch_off_the_extension_cord/episode_6/cam_high/frame_000000.jpg", "images/droid_oxe/Switch_off_the_extension_cord/episode_6/cam_high/frame_000188.jpg", "images/droid_oxe/Switch_off_the_extension_cord/episode_8/cam_high/frame_000180.jpg", "images/droid_oxe/Switch_off_the_extension_cord/episode_8/cam_left_wrist/frame_000180.jpg", "images/droid_oxe/Switch_off_the_extension_cord/episode_8/cam_right_wrist/frame_000180.jpg", "images/droid_oxe/Switch_off_the_extension_cord/episode_8/cam_high/frame_000075.jpg", "images/droid_oxe/Switch_off_the_extension_cord/episode_8/cam_left_wrist/frame_000075.jpg", "images/droid_oxe/Switch_off_the_extension_cord/episode_8/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Switch off the extension cord`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Switch off the extension cord`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -58.3, "hop_absolute_value": 58.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Remove_the_yellow_object_from_the_wooden_structure_episode_0-train-bf_000105-af_000030-minus-0000030", "task": "Remove the yellow object from the wooden structure", "image": [ "images/droid_oxe/Remove_the_yellow_object_from_the_wooden_structure/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Remove_the_yellow_object_from_the_wooden_structure/episode_0/cam_high/frame_000195.jpg", "images/droid_oxe/Remove_the_yellow_object_from_the_wooden_structure/episode_0/cam_high/frame_000105.jpg", "images/droid_oxe/Remove_the_yellow_object_from_the_wooden_structure/episode_0/cam_left_wrist/frame_000105.jpg", "images/droid_oxe/Remove_the_yellow_object_from_the_wooden_structure/episode_0/cam_right_wrist/frame_000105.jpg", "images/droid_oxe/Remove_the_yellow_object_from_the_wooden_structure/episode_0/cam_high/frame_000030.jpg", "images/droid_oxe/Remove_the_yellow_object_from_the_wooden_structure/episode_0/cam_left_wrist/frame_000030.jpg", "images/droid_oxe/Remove_the_yellow_object_from_the_wooden_structure/episode_0/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the yellow object from the wooden structure`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the yellow object from the wooden structure`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Wipe_the_desk_with_the_white_towel_episode_0-train-bf_000510-af_000060-minus-0000067", "task": "Wipe the desk with the white towel", "image": [ "images/droid_oxe/Wipe_the_desk_with_the_white_towel/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Wipe_the_desk_with_the_white_towel/episode_0/cam_high/frame_000557.jpg", "images/droid_oxe/Wipe_the_desk_with_the_white_towel/episode_0/cam_high/frame_000510.jpg", "images/droid_oxe/Wipe_the_desk_with_the_white_towel/episode_0/cam_left_wrist/frame_000510.jpg", "images/droid_oxe/Wipe_the_desk_with_the_white_towel/episode_0/cam_right_wrist/frame_000510.jpg", "images/droid_oxe/Wipe_the_desk_with_the_white_towel/episode_0/cam_high/frame_000060.jpg", "images/droid_oxe/Wipe_the_desk_with_the_white_towel/episode_0/cam_left_wrist/frame_000060.jpg", "images/droid_oxe/Wipe_the_desk_with_the_white_towel/episode_0/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wipe the desk with the white towel`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wipe the desk with the white towel`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -88.2, "hop_absolute_value": 88.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Put_the_pot_on_the_cooker_and_cover_it_with_the_lid_episode_0-train-bf_000810-af_000720-minus-0000711", "task": "Put the pot on the cooker and cover it with the lid", "image": [ "images/droid_oxe/Put_the_pot_on_the_cooker_and_cover_it_with_the_lid/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Put_the_pot_on_the_cooker_and_cover_it_with_the_lid/episode_0/cam_high/frame_000882.jpg", "images/droid_oxe/Put_the_pot_on_the_cooker_and_cover_it_with_the_lid/episode_0/cam_high/frame_000810.jpg", "images/droid_oxe/Put_the_pot_on_the_cooker_and_cover_it_with_the_lid/episode_0/cam_left_wrist/frame_000810.jpg", "images/droid_oxe/Put_the_pot_on_the_cooker_and_cover_it_with_the_lid/episode_0/cam_right_wrist/frame_000810.jpg", "images/droid_oxe/Put_the_pot_on_the_cooker_and_cover_it_with_the_lid/episode_0/cam_high/frame_000720.jpg", "images/droid_oxe/Put_the_pot_on_the_cooker_and_cover_it_with_the_lid/episode_0/cam_left_wrist/frame_000720.jpg", "images/droid_oxe/Put_the_pot_on_the_cooker_and_cover_it_with_the_lid/episode_0/cam_right_wrist/frame_000720.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the pot on the cooker and cover it with the lid`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the pot on the cooker and cover it with the lid`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "droid_oxe_Stir_the_contents_in_the_green_bow_using_the_spatula_episode_0-train-bf_000270-af_000195-minus-0000692", "task": "Stir the contents in the green bow using the spatula", "image": [ "images/droid_oxe/Stir_the_contents_in_the_green_bow_using_the_spatula/episode_0/cam_high/frame_000000.jpg", "images/droid_oxe/Stir_the_contents_in_the_green_bow_using_the_spatula/episode_0/cam_high/frame_000482.jpg", "images/droid_oxe/Stir_the_contents_in_the_green_bow_using_the_spatula/episode_0/cam_high/frame_000270.jpg", "images/droid_oxe/Stir_the_contents_in_the_green_bow_using_the_spatula/episode_0/cam_left_wrist/frame_000270.jpg", "images/droid_oxe/Stir_the_contents_in_the_green_bow_using_the_spatula/episode_0/cam_right_wrist/frame_000270.jpg", "images/droid_oxe/Stir_the_contents_in_the_green_bow_using_the_spatula/episode_0/cam_high/frame_000195.jpg", "images/droid_oxe/Stir_the_contents_in_the_green_bow_using_the_spatula/episode_0/cam_left_wrist/frame_000195.jpg", "images/droid_oxe/Stir_the_contents_in_the_green_bow_using_the_spatula/episode_0/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Stir the contents in the green bow using the spatula`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Stir the contents in the green bow using the spatula`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -27.8, "hop_absolute_value": 27.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "droid_oxe" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_20-train-bf_000435-af_000915-plus-0000474", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_38/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_38/cam_high/frame_001151.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_20/cam_high/frame_000435.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_20/cam_left_wrist/frame_000435.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_20/cam_right_wrist/frame_000435.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_20/cam_high/frame_000915.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_20/cam_left_wrist/frame_000915.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_20/cam_right_wrist/frame_000915.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 69.6, "hop_absolute_value": 69.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Place_the_paper_cup_on_the_table_episode_16-train-bf_000120-af_000135-plus-0000074", "task": "Place the paper cup on the table", "image": [ "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_19/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_19/cam_high/frame_000147.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_16/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_16/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_16/cam_right_wrist/frame_000120.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_16/cam_high/frame_000135.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_16/cam_left_wrist/frame_000135.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_16/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the paper cup on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the paper cup on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_26-train-bf_000000-af_001425-plus-0000516", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_2/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_2/cam_high/frame_001283.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_26/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_26/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_26/cam_right_wrist/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_26/cam_high/frame_001425.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_26/cam_left_wrist/frame_001425.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_26/cam_right_wrist/frame_001425.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.4, "hop_absolute_value": 70.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_nightstand_drawer_episode_8-train-bf_000015-af_000045-plus-0000026", "task": "Close the nightstand drawer", "image": [ "images/galaxea_r1lite/Close_the_nightstand_drawer/episode_19/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_nightstand_drawer/episode_19/cam_high/frame_000111.jpg", "images/galaxea_r1lite/Close_the_nightstand_drawer/episode_8/cam_high/frame_000015.jpg", "images/galaxea_r1lite/Close_the_nightstand_drawer/episode_8/cam_left_wrist/frame_000015.jpg", "images/galaxea_r1lite/Close_the_nightstand_drawer/episode_8/cam_right_wrist/frame_000015.jpg", "images/galaxea_r1lite/Close_the_nightstand_drawer/episode_8/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Close_the_nightstand_drawer/episode_8/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Close_the_nightstand_drawer/episode_8/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the nightstand drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the nightstand drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_microwave_oven_episode_3-train-bf_000000-af_000165-plus-0000312", "task": "Open the microwave oven", "image": [ "images/galaxea_r1lite/Open_the_microwave_oven/episode_5/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_5/cam_high/frame_000941.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_3/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_3/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_3/cam_right_wrist/frame_000000.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_3/cam_high/frame_000165.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_3/cam_left_wrist/frame_000165.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_3/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the microwave oven`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the microwave oven`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 42.3, "hop_absolute_value": 42.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_41-train-bf_000000-af_000240-plus-0000340", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_24/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_24/cam_high/frame_000588.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_41/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_41/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_41/cam_right_wrist/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_41/cam_high/frame_000240.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_41/cam_left_wrist/frame_000240.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_41/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 26.2, "hop_absolute_value": 26.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_shoes_into_the_shoe_box_episode_2-train-bf_000015-af_000105-plus-0000105", "task": "Put the shoes into the shoe box", "image": [ "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_18/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_18/cam_high/frame_000109.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_2/cam_high/frame_000015.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_2/cam_left_wrist/frame_000015.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_2/cam_right_wrist/frame_000015.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_2/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_2/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_2/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the shoes into the shoe box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the shoes into the shoe box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_14-train-bf_000315-af_001245-plus-0000462", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_21/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_21/cam_high/frame_001008.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_14/cam_high/frame_000315.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_14/cam_left_wrist/frame_000315.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_14/cam_right_wrist/frame_000315.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_14/cam_high/frame_001245.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_14/cam_left_wrist/frame_001245.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_14/cam_right_wrist/frame_001245.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.2, "hop_absolute_value": 60.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_42-train-bf_000360-af_000405-plus-0000872", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_42/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_42/cam_high/frame_001755.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_42/cam_high/frame_000360.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_42/cam_left_wrist/frame_000360.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_42/cam_right_wrist/frame_000360.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_42/cam_high/frame_000405.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_42/cam_left_wrist/frame_000405.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_42/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.2, "hop_absolute_value": 3.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_slide_door_episode_3-train-bf_000000-af_000060-plus-0000848", "task": "Close the slide door", "image": [ "images/galaxea_r1lite/Close_the_slide_door/episode_0/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_0/cam_high/frame_000761.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_3/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_3/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_3/cam_right_wrist/frame_000000.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_3/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_3/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_3/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 7.5, "hop_absolute_value": 7.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Take_the_storage_box_out_of_the_drawer_and_put_it_on_the_table_episode_17-train-bf_000000-af_000030-plus-0000144", "task": "Take the storage box out of the drawer and put it on the table", "image": [ "images/galaxea_r1lite/Take_the_storage_box_out_of_the_drawer_and_put_it_on_the_table/episode_17/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_storage_box_out_of_the_drawer_and_put_it_on_the_table/episode_17/cam_high/frame_000240.jpg", "images/galaxea_r1lite/Take_the_storage_box_out_of_the_drawer_and_put_it_on_the_table/episode_17/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_storage_box_out_of_the_drawer_and_put_it_on_the_table/episode_17/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Take_the_storage_box_out_of_the_drawer_and_put_it_on_the_table/episode_17/cam_right_wrist/frame_000000.jpg", "images/galaxea_r1lite/Take_the_storage_box_out_of_the_drawer_and_put_it_on_the_table/episode_17/cam_high/frame_000030.jpg", "images/galaxea_r1lite/Take_the_storage_box_out_of_the_drawer_and_put_it_on_the_table/episode_17/cam_left_wrist/frame_000030.jpg", "images/galaxea_r1lite/Take_the_storage_box_out_of_the_drawer_and_put_it_on_the_table/episode_17/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the storage box out of the drawer and put it on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the storage box out of the drawer and put it on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_1-train-bf_000930-af_000945-plus-0000866", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_39/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_39/cam_high/frame_001217.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_1/cam_high/frame_000930.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_1/cam_left_wrist/frame_000930.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_1/cam_right_wrist/frame_000930.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_1/cam_high/frame_000945.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_1/cam_left_wrist/frame_000945.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_1/cam_right_wrist/frame_000945.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.2, "hop_absolute_value": 2.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one_episode_9-train-bf_000405-af_000420-plus-0000643", "task": "Put the remote controls on the table in the drawer one by one", "image": [ "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_9/cam_high/frame_000567.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_9/cam_high/frame_000405.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_9/cam_left_wrist/frame_000405.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_9/cam_right_wrist/frame_000405.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_9/cam_high/frame_000420.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_9/cam_left_wrist/frame_000420.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_9/cam_right_wrist/frame_000420.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the remote controls on the table in the drawer one by one`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the remote controls on the table in the drawer one by one`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_53-train-bf_000210-af_000705-plus-0000393", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_high/frame_001568.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_high/frame_000210.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_left_wrist/frame_000210.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_right_wrist/frame_000210.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_high/frame_000705.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_left_wrist/frame_000705.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_right_wrist/frame_000705.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 36.7, "hop_absolute_value": 36.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Push_the_blue_chair_back_under_the_dining_table_episode_10-train-bf_000165-af_000255-plus-0000305", "task": "Push the blue chair back under the dining table", "image": [ "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_2/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_2/cam_high/frame_000492.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_high/frame_000165.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_left_wrist/frame_000165.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_right_wrist/frame_000165.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_high/frame_000255.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_left_wrist/frame_000255.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the blue chair back under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the blue chair back under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_leftover_food_episode_4-train-bf_000915-af_001065-plus-0000439", "task": "Dispose of leftover food", "image": [ "images/galaxea_r1lite/Dispose_of_leftover_food/episode_13/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_13/cam_high/frame_000951.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_4/cam_high/frame_000915.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_4/cam_left_wrist/frame_000915.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_4/cam_right_wrist/frame_000915.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_4/cam_high/frame_001065.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_4/cam_left_wrist/frame_001065.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_4/cam_right_wrist/frame_001065.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of leftover food`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of leftover food`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_33-train-bf_000180-af_000885-plus-0000439", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_42/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_42/cam_high/frame_001755.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_33/cam_high/frame_000180.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_33/cam_left_wrist/frame_000180.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_33/cam_right_wrist/frame_000180.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_33/cam_high/frame_000885.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_33/cam_left_wrist/frame_000885.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_33/cam_right_wrist/frame_000885.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 58.0, "hop_absolute_value": 58.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Push_the_blue_chair_back_under_the_dining_table_episode_13-train-bf_000000-af_000120-plus-0000260", "task": "Push the blue chair back under the dining table", "image": [ "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_high/frame_000439.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_13/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_13/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_13/cam_right_wrist/frame_000000.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_13/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_13/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_13/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the blue chair back under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the blue chair back under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 34.8, "hop_absolute_value": 34.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_microwave_oven_episode_17-train-bf_000045-af_000165-plus-0000800", "task": "Open the microwave oven", "image": [ "images/galaxea_r1lite/Open_the_microwave_oven/episode_1/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_1/cam_high/frame_000305.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_17/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_17/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_17/cam_right_wrist/frame_000045.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_17/cam_high/frame_000165.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_17/cam_left_wrist/frame_000165.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_17/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the microwave oven`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the microwave oven`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_28-train-bf_000450-af_000585-plus-0000328", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_23/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_23/cam_high/frame_000882.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_28/cam_high/frame_000450.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_28/cam_left_wrist/frame_000450.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_28/cam_right_wrist/frame_000450.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_28/cam_high/frame_000585.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_28/cam_left_wrist/frame_000585.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_28/cam_right_wrist/frame_000585.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 27.3, "hop_absolute_value": 27.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_gray_chair_out_from_under_the_dining_table_episode_14-train-bf_000165-af_000217-plus-0000194", "task": "Pull the gray chair out from under the dining table", "image": [ "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_1/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_1/cam_high/frame_000270.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_14/cam_high/frame_000165.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_14/cam_left_wrist/frame_000165.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_14/cam_right_wrist/frame_000165.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_14/cam_high/frame_000217.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_14/cam_left_wrist/frame_000217.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_14/cam_right_wrist/frame_000217.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the gray chair out from under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the gray chair out from under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Push_the_gray_chair_back_under_the_dining_table_episode_7-train-bf_000030-af_000255-plus-0000385", "task": "Push the gray chair back under the dining table", "image": [ "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_10/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_10/cam_high/frame_000380.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_7/cam_high/frame_000030.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_7/cam_left_wrist/frame_000030.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_7/cam_right_wrist/frame_000030.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_7/cam_high/frame_000255.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_7/cam_left_wrist/frame_000255.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_7/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the gray chair back under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the gray chair back under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_slide_door_episode_18-train-bf_000015-af_000375-plus-0000492", "task": "Open the slide door", "image": [ "images/galaxea_r1lite/Open_the_slide_door/episode_4/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_4/cam_high/frame_000570.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_18/cam_high/frame_000015.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_18/cam_left_wrist/frame_000015.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_18/cam_right_wrist/frame_000015.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_18/cam_high/frame_000375.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_18/cam_left_wrist/frame_000375.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_18/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 92.3, "hop_absolute_value": 92.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_38-train-bf_001095-af_001197-plus-0000577", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_38/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_38/cam_high/frame_001197.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_38/cam_high/frame_001095.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_38/cam_left_wrist/frame_001095.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_38/cam_right_wrist/frame_001095.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_38/cam_high/frame_001197.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_38/cam_left_wrist/frame_001197.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_38/cam_right_wrist/frame_001197.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_leftover_food_episode_18-train-bf_000630-af_000855-plus-0000500", "task": "Dispose of leftover food", "image": [ "images/galaxea_r1lite/Dispose_of_leftover_food/episode_3/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_3/cam_high/frame_001142.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_18/cam_high/frame_000630.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_18/cam_left_wrist/frame_000630.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_18/cam_right_wrist/frame_000630.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_18/cam_high/frame_000855.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_18/cam_left_wrist/frame_000855.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_18/cam_right_wrist/frame_000855.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of leftover food`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of leftover food`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 78.9, "hop_absolute_value": 78.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_blue_chair_out_from_under_the_dining_table_episode_9-train-bf_000195-af_000225-plus-0000263", "task": "Pull the blue chair out from under the dining table", "image": [ "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_16/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_16/cam_high/frame_000372.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_9/cam_high/frame_000195.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_9/cam_left_wrist/frame_000195.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_9/cam_right_wrist/frame_000195.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_9/cam_high/frame_000225.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_9/cam_left_wrist/frame_000225.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_9/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the blue chair out from under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the blue chair out from under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 15.4, "hop_absolute_value": 15.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_inner_curtains_open_on_both_sides_episode_5-train-bf_000180-af_000315-plus-0000359", "task": "Pull the inner curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_19/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_19/cam_high/frame_000523.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_high/frame_000180.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_left_wrist/frame_000180.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_right_wrist/frame_000180.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_high/frame_000315.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_left_wrist/frame_000315.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_right_wrist/frame_000315.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the inner curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the inner curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 56.2, "hop_absolute_value": 56.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_outer_curtains_on_both_sides_episode_0-train-bf_000330-af_000360-plus-0000351", "task": "Close the outer curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_7/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_7/cam_high/frame_000555.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_0/cam_high/frame_000330.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_0/cam_left_wrist/frame_000330.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_0/cam_right_wrist/frame_000330.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_0/cam_high/frame_000360.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_0/cam_left_wrist/frame_000360.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_0/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the outer curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the outer curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Push_the_gray_chair_back_under_the_dining_table_episode_14-train-bf_000195-af_000240-plus-0000248", "task": "Push the gray chair back under the dining table", "image": [ "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_2/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_2/cam_high/frame_000331.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_14/cam_high/frame_000195.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_14/cam_left_wrist/frame_000195.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_14/cam_right_wrist/frame_000195.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_14/cam_high/frame_000240.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_14/cam_left_wrist/frame_000240.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_14/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the gray chair back under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the gray chair back under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 37.5, "hop_absolute_value": 37.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_outer_curtains_on_both_sides_episode_12-train-bf_000315-af_000375-plus-0000856", "task": "Close the outer curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_14/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_14/cam_high/frame_000577.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_12/cam_high/frame_000315.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_12/cam_left_wrist/frame_000315.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_12/cam_right_wrist/frame_000315.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_12/cam_high/frame_000375.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_12/cam_left_wrist/frame_000375.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_12/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the outer curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the outer curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 23.5, "hop_absolute_value": 23.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_51-train-bf_000165-af_000990-plus-0000524", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_51/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_51/cam_high/frame_001197.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_51/cam_high/frame_000165.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_51/cam_left_wrist/frame_000165.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_51/cam_right_wrist/frame_000165.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_51/cam_high/frame_000990.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_51/cam_left_wrist/frame_000990.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_51/cam_right_wrist/frame_000990.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 80.9, "hop_absolute_value": 80.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_shorts_episode_10-train-bf_000060-af_000150-plus-0000907", "task": "Unfold the shorts", "image": [ "images/galaxea_r1lite/Unfold_the_shorts/episode_24/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_24/cam_high/frame_002149.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_10/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_10/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_10/cam_right_wrist/frame_000060.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_10/cam_high/frame_000150.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_10/cam_left_wrist/frame_000150.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_10/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 13.6, "hop_absolute_value": 13.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Take_the_yoghurt_out_of_the_freezer_episode_6-train-bf_000240-af_000285-plus-0000710", "task": "Take the yoghurt out of the freezer", "image": [ "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_21/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_21/cam_high/frame_000186.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_6/cam_high/frame_000240.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_6/cam_left_wrist/frame_000240.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_6/cam_right_wrist/frame_000240.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_6/cam_high/frame_000285.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_6/cam_left_wrist/frame_000285.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_6/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the yoghurt out of the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the yoghurt out of the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_garbage_in_the_trash_can_episode_1-train-bf_001545-af_001605-plus-0000344", "task": "Dispose of garbage in the trash can", "image": [ "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_24/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_24/cam_high/frame_001178.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_1/cam_high/frame_001545.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_1/cam_left_wrist/frame_001545.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_1/cam_right_wrist/frame_001545.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_1/cam_high/frame_001605.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_1/cam_left_wrist/frame_001605.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_1/cam_right_wrist/frame_001605.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of garbage in the trash can`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of garbage in the trash can`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 23.5, "hop_absolute_value": 23.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Place_the_ceramic_cup_on_the_table_episode_3-train-bf_000105-af_000225-plus-0000259", "task": "Place the ceramic cup on the table", "image": [ "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_3/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_3/cam_high/frame_000282.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_3/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_3/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_3/cam_right_wrist/frame_000105.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_3/cam_high/frame_000225.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_3/cam_left_wrist/frame_000225.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_3/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the ceramic cup on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the ceramic cup on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 72.7, "hop_absolute_value": 72.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_outer_curtains_open_on_both_sides_episode_1-train-bf_000240-af_000278-plus-0000300", "task": "Pull the outer curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_high/frame_000614.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_1/cam_high/frame_000240.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_1/cam_left_wrist/frame_000240.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_1/cam_right_wrist/frame_000240.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_1/cam_high/frame_000278.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_1/cam_left_wrist/frame_000278.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_1/cam_right_wrist/frame_000278.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the outer curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the outer curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_drawer_episode_23-train-bf_000000-af_000300-plus-0000405", "task": "Open the drawer", "image": [ "images/galaxea_r1lite/Open_the_drawer/episode_17/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_17/cam_high/frame_000245.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_23/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_23/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_23/cam_right_wrist/frame_000000.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_23/cam_high/frame_000300.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_23/cam_left_wrist/frame_000300.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_23/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 55.6, "hop_absolute_value": 55.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_inner_curtains_on_both_sides_episode_9-train-bf_000000-af_000360-plus-0000472", "task": "Close the inner curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_15/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_15/cam_high/frame_000398.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_9/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_9/cam_right_wrist/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_9/cam_high/frame_000360.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_9/cam_left_wrist/frame_000360.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_9/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the inner curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the inner curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 77.4, "hop_absolute_value": 77.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_slide_door_episode_7-train-bf_000060-af_000630-plus-0000480", "task": "Open the slide door", "image": [ "images/galaxea_r1lite/Open_the_slide_door/episode_15/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_15/cam_high/frame_000675.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_7/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_7/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_7/cam_right_wrist/frame_000060.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_7/cam_high/frame_000630.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_7/cam_left_wrist/frame_000630.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_7/cam_right_wrist/frame_000630.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 73.1, "hop_absolute_value": 73.1, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_plate_in_the_microwave_oven_episode_20-train-bf_000195-af_000263-plus-0000277", "task": "Put the plate in the microwave oven", "image": [ "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_24/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_24/cam_high/frame_000269.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_20/cam_high/frame_000195.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_20/cam_left_wrist/frame_000195.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_20/cam_right_wrist/frame_000195.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_20/cam_high/frame_000263.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_20/cam_left_wrist/frame_000263.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_20/cam_right_wrist/frame_000263.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the plate in the microwave oven`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the plate in the microwave oven`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_43-train-bf_000030-af_000060-plus-0000706", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_46/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_46/cam_high/frame_001029.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_43/cam_high/frame_000030.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_43/cam_left_wrist/frame_000030.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_43/cam_right_wrist/frame_000030.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_43/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_43/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_43/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.9, "hop_absolute_value": 2.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_gray_chair_out_from_under_the_dining_table_episode_17-train-bf_000210-af_000420-plus-0000477", "task": "Pull the gray chair out from under the dining table", "image": [ "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_high/frame_000460.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_high/frame_000210.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_left_wrist/frame_000210.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_right_wrist/frame_000210.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_high/frame_000420.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_left_wrist/frame_000420.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_right_wrist/frame_000420.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the gray chair out from under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the gray chair out from under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_23-train-bf_000075-af_000135-plus-0000902", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_44/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_44/cam_high/frame_000983.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_23/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_23/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_23/cam_right_wrist/frame_000075.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_23/cam_high/frame_000135.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_23/cam_left_wrist/frame_000135.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_23/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 7.5, "hop_absolute_value": 7.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_yoghurt_in_the_freezer_episode_21-train-bf_000105-af_000180-plus-0000221", "task": "Put the yoghurt in the freezer", "image": [ "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_16/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_16/cam_high/frame_000281.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_21/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_21/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_21/cam_right_wrist/frame_000105.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_21/cam_high/frame_000180.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_21/cam_left_wrist/frame_000180.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_21/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the yoghurt in the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the yoghurt in the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 45.5, "hop_absolute_value": 45.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_57-train-bf_000330-af_000720-plus-0000360", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_17/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_17/cam_high/frame_001729.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_57/cam_high/frame_000330.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_57/cam_left_wrist/frame_000330.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_57/cam_right_wrist/frame_000330.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_57/cam_high/frame_000720.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_57/cam_left_wrist/frame_000720.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_57/cam_right_wrist/frame_000720.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 29.2, "hop_absolute_value": 29.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_microwave_oven_episode_1-train-bf_000075-af_000285-plus-0000362", "task": "Open the microwave oven", "image": [ "images/galaxea_r1lite/Open_the_microwave_oven/episode_22/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_22/cam_high/frame_000159.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_1/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_1/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_1/cam_right_wrist/frame_000075.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_1/cam_high/frame_000285.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_1/cam_left_wrist/frame_000285.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_1/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the microwave oven`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the microwave oven`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 93.3, "hop_absolute_value": 93.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_blue_chair_out_from_under_the_dining_table_episode_24-train-bf_000180-af_000270-plus-0000640", "task": "Pull the blue chair out from under the dining table", "image": [ "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_19/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_19/cam_high/frame_000482.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_24/cam_high/frame_000180.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_24/cam_left_wrist/frame_000180.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_24/cam_right_wrist/frame_000180.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_24/cam_high/frame_000270.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_24/cam_left_wrist/frame_000270.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_24/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the blue chair out from under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the blue chair out from under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 42.9, "hop_absolute_value": 42.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_inner_curtains_on_both_sides_episode_24-train-bf_000000-af_000255-plus-0000367", "task": "Close the inner curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_12/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_12/cam_high/frame_000310.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_24/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_24/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_24/cam_right_wrist/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_24/cam_high/frame_000255.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_24/cam_left_wrist/frame_000255.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_24/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the inner curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the inner curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 77.3, "hop_absolute_value": 77.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_15-train-bf_000420-af_000885-plus-0000405", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_52/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_52/cam_high/frame_001259.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_15/cam_high/frame_000420.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_15/cam_left_wrist/frame_000420.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_15/cam_right_wrist/frame_000420.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_15/cam_high/frame_000885.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_15/cam_left_wrist/frame_000885.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_15/cam_right_wrist/frame_000885.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 46.3, "hop_absolute_value": 46.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_10-train-bf_000690-af_001020-plus-0000512", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_10/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_10/cam_high/frame_001106.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_10/cam_high/frame_000690.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_10/cam_left_wrist/frame_000690.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_10/cam_right_wrist/frame_000690.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_10/cam_high/frame_001020.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_10/cam_left_wrist/frame_001020.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_10/cam_right_wrist/frame_001020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 81.5, "hop_absolute_value": 81.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Push_the_gray_chair_back_under_the_dining_table_episode_0-train-bf_000000-af_000300-plus-0000395", "task": "Push the gray chair back under the dining table", "image": [ "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_0/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_0/cam_high/frame_000326.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_0/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_0/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_0/cam_right_wrist/frame_000000.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_0/cam_high/frame_000300.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_0/cam_left_wrist/frame_000300.jpg", "images/galaxea_r1lite/Push_the_gray_chair_back_under_the_dining_table/episode_0/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the gray chair back under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the gray chair back under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 95.2, "hop_absolute_value": 95.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_shorts_episode_12-train-bf_000210-af_000630-plus-0000576", "task": "Unfold the shorts", "image": [ "images/galaxea_r1lite/Unfold_the_shorts/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_9/cam_high/frame_001907.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_12/cam_high/frame_000210.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_12/cam_left_wrist/frame_000210.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_12/cam_right_wrist/frame_000210.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_12/cam_high/frame_000630.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_12/cam_left_wrist/frame_000630.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_12/cam_right_wrist/frame_000630.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 96.6, "hop_absolute_value": 96.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_40-train-bf_000120-af_000750-plus-0000515", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_40/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_40/cam_high/frame_000918.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_40/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_40/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_40/cam_right_wrist/frame_000120.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_40/cam_high/frame_000750.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_40/cam_left_wrist/frame_000750.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_40/cam_right_wrist/frame_000750.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 79.2, "hop_absolute_value": 79.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_drawer_episode_23-train-bf_000105-af_000390-plus-0000428", "task": "Open the drawer", "image": [ "images/galaxea_r1lite/Open_the_drawer/episode_5/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_5/cam_high/frame_000270.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_23/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_23/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_23/cam_right_wrist/frame_000105.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_23/cam_high/frame_000390.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_23/cam_left_wrist/frame_000390.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_23/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 65.5, "hop_absolute_value": 65.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_paper_cup_in_the_freezer_episode_15-train-bf_000045-af_000255-plus-0000317", "task": "Put the paper cup in the freezer", "image": [ "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_6/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_6/cam_high/frame_000474.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_15/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_15/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_15/cam_right_wrist/frame_000045.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_15/cam_high/frame_000255.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_15/cam_left_wrist/frame_000255.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_15/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the paper cup in the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the paper cup in the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_sweatshirts_episode_5-train-bf_000720-af_000735-plus-0000916", "task": "Unfold the sweatshirts", "image": [ "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_3/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_3/cam_high/frame_001277.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_5/cam_high/frame_000720.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_5/cam_left_wrist/frame_000720.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_5/cam_right_wrist/frame_000720.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_5/cam_high/frame_000735.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_5/cam_left_wrist/frame_000735.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_5/cam_right_wrist/frame_000735.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the sweatshirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the sweatshirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.8, "hop_absolute_value": 1.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_microwave_oven_episode_20-train-bf_000075-af_000105-plus-0000148", "task": "Open the microwave oven", "image": [ "images/galaxea_r1lite/Open_the_microwave_oven/episode_20/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_20/cam_high/frame_000249.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_20/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_20/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_20/cam_right_wrist/frame_000075.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_20/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_20/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_20/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the microwave oven`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the microwave oven`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 18.2, "hop_absolute_value": 18.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_garbage_in_the_trash_can_episode_18-train-bf_000540-af_000555-plus-0000364", "task": "Dispose of garbage in the trash can", "image": [ "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_17/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_17/cam_high/frame_000956.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_18/cam_high/frame_000540.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_18/cam_left_wrist/frame_000540.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_18/cam_right_wrist/frame_000540.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_18/cam_high/frame_000555.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_18/cam_left_wrist/frame_000555.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_18/cam_right_wrist/frame_000555.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of garbage in the trash can`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of garbage in the trash can`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_shoe_box_episode_3-train-bf_000075-af_000210-plus-0000216", "task": "Open the shoe box", "image": [ "images/galaxea_r1lite/Open_the_shoe_box/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_9/cam_high/frame_000250.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_3/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_3/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_3/cam_right_wrist/frame_000075.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_3/cam_high/frame_000210.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_3/cam_left_wrist/frame_000210.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_3/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the shoe box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the shoe box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 90.0, "hop_absolute_value": 90.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Take_the_paper_cup_out_of_the_freezer_episode_13-train-bf_000045-af_000161-plus-0000105", "task": "Take the paper cup out of the freezer", "image": [ "images/galaxea_r1lite/Take_the_paper_cup_out_of_the_freezer/episode_13/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_paper_cup_out_of_the_freezer/episode_13/cam_high/frame_000161.jpg", "images/galaxea_r1lite/Take_the_paper_cup_out_of_the_freezer/episode_13/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Take_the_paper_cup_out_of_the_freezer/episode_13/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Take_the_paper_cup_out_of_the_freezer/episode_13/cam_right_wrist/frame_000045.jpg", "images/galaxea_r1lite/Take_the_paper_cup_out_of_the_freezer/episode_13/cam_high/frame_000161.jpg", "images/galaxea_r1lite/Take_the_paper_cup_out_of_the_freezer/episode_13/cam_left_wrist/frame_000161.jpg", "images/galaxea_r1lite/Take_the_paper_cup_out_of_the_freezer/episode_13/cam_right_wrist/frame_000161.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the paper cup out of the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the paper cup out of the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_slide_door_episode_19-train-bf_000105-af_000360-plus-0000779", "task": "Open the slide door", "image": [ "images/galaxea_r1lite/Open_the_slide_door/episode_0/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_0/cam_high/frame_000832.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_19/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_19/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_19/cam_right_wrist/frame_000105.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_19/cam_high/frame_000360.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_19/cam_left_wrist/frame_000360.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_19/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 81.0, "hop_absolute_value": 81.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_garbage_in_the_trash_can_episode_17-train-bf_000120-af_000600-plus-0000426", "task": "Dispose of garbage in the trash can", "image": [ "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_12/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_12/cam_high/frame_001973.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_17/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_17/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_17/cam_right_wrist/frame_000120.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_17/cam_high/frame_000600.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_17/cam_left_wrist/frame_000600.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_17/cam_right_wrist/frame_000600.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of garbage in the trash can`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of garbage in the trash can`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 58.2, "hop_absolute_value": 58.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_0-train-bf_000045-af_000075-plus-0000759", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_38/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_38/cam_high/frame_001197.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_0/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_0/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_0/cam_right_wrist/frame_000045.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_0/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_0/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_0/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.7, "hop_absolute_value": 1.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_slide_door_episode_19-train-bf_000105-af_000240-plus-0000320", "task": "Close the slide door", "image": [ "images/galaxea_r1lite/Close_the_slide_door/episode_12/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_12/cam_high/frame_001039.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_19/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_19/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_19/cam_right_wrist/frame_000105.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_19/cam_high/frame_000240.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_19/cam_left_wrist/frame_000240.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_19/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one_episode_14-train-bf_000120-af_000135-plus-0000104", "task": "Put the remote controls on the table in the drawer one by one", "image": [ "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_6/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_6/cam_high/frame_000443.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_14/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_14/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_14/cam_right_wrist/frame_000120.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_14/cam_high/frame_000135.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_14/cam_left_wrist/frame_000135.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_14/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the remote controls on the table in the drawer one by one`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the remote controls on the table in the drawer one by one`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_shorts_episode_5-train-bf_000030-af_000045-plus-0000676", "task": "Fold the shorts", "image": [ "images/galaxea_r1lite/Fold_the_shorts/episode_25/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_25/cam_high/frame_000557.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_5/cam_high/frame_000030.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_5/cam_left_wrist/frame_000030.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_5/cam_right_wrist/frame_000030.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_5/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_5/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_5/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.7, "hop_absolute_value": 1.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_slide_door_episode_7-train-bf_000060-af_000645-plus-0000425", "task": "Close the slide door", "image": [ "images/galaxea_r1lite/Close_the_slide_door/episode_4/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_4/cam_high/frame_001010.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_7/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_7/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_7/cam_right_wrist/frame_000060.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_7/cam_high/frame_000645.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_7/cam_left_wrist/frame_000645.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_7/cam_right_wrist/frame_000645.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.4, "hop_absolute_value": 57.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_shorts_episode_26-train-bf_000075-af_000450-plus-0000418", "task": "Unfold the shorts", "image": [ "images/galaxea_r1lite/Unfold_the_shorts/episode_19/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_19/cam_high/frame_001807.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_26/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_26/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_26/cam_right_wrist/frame_000075.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_26/cam_high/frame_000450.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_26/cam_left_wrist/frame_000450.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_26/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 56.8, "hop_absolute_value": 56.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_drawer_episode_9-train-bf_000165-af_000345-plus-0000768", "task": "Open the drawer", "image": [ "images/galaxea_r1lite/Open_the_drawer/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_9/cam_high/frame_000426.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_9/cam_high/frame_000165.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_9/cam_left_wrist/frame_000165.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_9/cam_right_wrist/frame_000165.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_9/cam_high/frame_000345.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_9/cam_left_wrist/frame_000345.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_9/cam_right_wrist/frame_000345.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.6, "hop_absolute_value": 70.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_paper_cup_in_the_freezer_episode_22-train-bf_000195-af_000270-plus-0000299", "task": "Put the paper cup in the freezer", "image": [ "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_4/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_4/cam_high/frame_000545.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_22/cam_high/frame_000195.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_22/cam_left_wrist/frame_000195.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_22/cam_right_wrist/frame_000195.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_22/cam_high/frame_000270.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_22/cam_left_wrist/frame_000270.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_22/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the paper cup in the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the paper cup in the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 41.7, "hop_absolute_value": 41.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_microwave_oven_episode_9-train-bf_000180-af_000375-plus-0000434", "task": "Open the microwave oven", "image": [ "images/galaxea_r1lite/Open_the_microwave_oven/episode_8/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_8/cam_high/frame_000339.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_9/cam_high/frame_000180.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_9/cam_left_wrist/frame_000180.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_9/cam_right_wrist/frame_000180.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_9/cam_high/frame_000375.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_9/cam_left_wrist/frame_000375.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_9/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the microwave oven`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the microwave oven`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 68.4, "hop_absolute_value": 68.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_inner_curtains_on_both_sides_episode_23-train-bf_000270-af_000300-plus-0000297", "task": "Close the inner curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_16/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_16/cam_high/frame_000486.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_23/cam_high/frame_000270.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_23/cam_left_wrist/frame_000270.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_23/cam_right_wrist/frame_000270.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_23/cam_high/frame_000300.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_23/cam_left_wrist/frame_000300.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_23/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the inner curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the inner curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_slide_door_episode_19-train-bf_000090-af_000405-plus-0000493", "task": "Open the slide door", "image": [ "images/galaxea_r1lite/Open_the_slide_door/episode_15/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_15/cam_high/frame_000675.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_19/cam_high/frame_000090.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_19/cam_left_wrist/frame_000090.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_19/cam_right_wrist/frame_000090.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_19/cam_high/frame_000405.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_19/cam_left_wrist/frame_000405.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_19/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 95.5, "hop_absolute_value": 95.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_shorts_episode_0-train-bf_000705-af_000720-plus-0000636", "task": "Unfold the shorts", "image": [ "images/galaxea_r1lite/Unfold_the_shorts/episode_19/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_19/cam_high/frame_001807.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_0/cam_high/frame_000705.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_0/cam_left_wrist/frame_000705.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_0/cam_right_wrist/frame_000705.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_0/cam_high/frame_000720.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_0/cam_left_wrist/frame_000720.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_0/cam_right_wrist/frame_000720.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.2, "hop_absolute_value": 1.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_ceramic_cup_in_the_freezer_episode_5-train-bf_000135-af_000435-plus-0000462", "task": "Put the ceramic cup in the freezer", "image": [ "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_12/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_12/cam_high/frame_000545.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_5/cam_high/frame_000135.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_5/cam_left_wrist/frame_000135.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_5/cam_right_wrist/frame_000135.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_5/cam_high/frame_000435.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_5/cam_left_wrist/frame_000435.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_5/cam_right_wrist/frame_000435.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the ceramic cup in the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the ceramic cup in the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_drawer_episode_16-train-bf_000195-af_000240-plus-0000241", "task": "Open the drawer", "image": [ "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_high/frame_000287.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_16/cam_high/frame_000195.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_16/cam_left_wrist/frame_000195.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_16/cam_right_wrist/frame_000195.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_16/cam_high/frame_000240.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_16/cam_left_wrist/frame_000240.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_16/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_inner_curtains_on_both_sides_episode_3-train-bf_000225-af_000390-plus-0000446", "task": "Close the inner curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_high/frame_000375.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_3/cam_high/frame_000225.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_3/cam_left_wrist/frame_000225.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_3/cam_right_wrist/frame_000225.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_3/cam_high/frame_000390.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_3/cam_left_wrist/frame_000390.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_3/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the inner curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the inner curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 91.7, "hop_absolute_value": 91.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_sweatshirts_episode_12-train-bf_000150-af_000480-plus-0000393", "task": "Unfold the sweatshirts", "image": [ "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_11/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_11/cam_high/frame_001030.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_12/cam_high/frame_000150.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_12/cam_left_wrist/frame_000150.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_12/cam_right_wrist/frame_000150.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_12/cam_high/frame_000480.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_12/cam_left_wrist/frame_000480.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_12/cam_right_wrist/frame_000480.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the sweatshirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the sweatshirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 44.0, "hop_absolute_value": 44.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_slide_door_episode_1-train-bf_000495-af_000705-plus-0000430", "task": "Close the slide door", "image": [ "images/galaxea_r1lite/Close_the_slide_door/episode_1/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_1/cam_high/frame_000850.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_1/cam_high/frame_000495.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_1/cam_left_wrist/frame_000495.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_1/cam_right_wrist/frame_000495.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_1/cam_high/frame_000705.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_1/cam_left_wrist/frame_000705.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_1/cam_right_wrist/frame_000705.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.9, "hop_absolute_value": 60.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Take_the_plate_out_of_the_microwave_episode_19-train-bf_000030-af_000075-plus-0000051", "task": "Take the plate out of the microwave", "image": [ "images/galaxea_r1lite/Take_the_plate_out_of_the_microwave/episode_16/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_plate_out_of_the_microwave/episode_16/cam_high/frame_000100.jpg", "images/galaxea_r1lite/Take_the_plate_out_of_the_microwave/episode_19/cam_high/frame_000030.jpg", "images/galaxea_r1lite/Take_the_plate_out_of_the_microwave/episode_19/cam_left_wrist/frame_000030.jpg", "images/galaxea_r1lite/Take_the_plate_out_of_the_microwave/episode_19/cam_right_wrist/frame_000030.jpg", "images/galaxea_r1lite/Take_the_plate_out_of_the_microwave/episode_19/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Take_the_plate_out_of_the_microwave/episode_19/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Take_the_plate_out_of_the_microwave/episode_19/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the plate out of the microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the plate out of the microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_slide_door_episode_18-train-bf_000330-af_000345-plus-0000616", "task": "Open the slide door", "image": [ "images/galaxea_r1lite/Open_the_slide_door/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_9/cam_high/frame_000642.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_18/cam_high/frame_000330.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_18/cam_left_wrist/frame_000330.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_18/cam_right_wrist/frame_000330.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_18/cam_high/frame_000345.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_18/cam_left_wrist/frame_000345.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_18/cam_right_wrist/frame_000345.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_ceramic_cup_in_the_freezer_episode_3-train-bf_000090-af_000345-plus-0000430", "task": "Put the ceramic cup in the freezer", "image": [ "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_6/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_6/cam_high/frame_000647.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_3/cam_high/frame_000090.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_3/cam_left_wrist/frame_000090.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_3/cam_right_wrist/frame_000090.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_3/cam_high/frame_000345.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_3/cam_left_wrist/frame_000345.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_3/cam_right_wrist/frame_000345.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the ceramic cup in the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the ceramic cup in the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.8, "hop_absolute_value": 70.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_slide_door_episode_18-train-bf_000270-af_000330-plus-0000331", "task": "Close the slide door", "image": [ "images/galaxea_r1lite/Close_the_slide_door/episode_8/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_8/cam_high/frame_000880.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_18/cam_high/frame_000270.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_18/cam_left_wrist/frame_000270.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_18/cam_right_wrist/frame_000270.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_18/cam_high/frame_000330.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_18/cam_left_wrist/frame_000330.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_18/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_53-train-bf_001155-af_001170-plus-0000946", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_54/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_54/cam_high/frame_001635.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_high/frame_001155.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_left_wrist/frame_001155.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_right_wrist/frame_001155.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_high/frame_001170.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_left_wrist/frame_001170.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_53/cam_right_wrist/frame_001170.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.7, "hop_absolute_value": 3.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_inner_curtains_open_on_both_sides_episode_3-train-bf_000060-af_000270-plus-0000378", "task": "Pull the inner curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_18/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_18/cam_high/frame_000312.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_3/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_3/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_3/cam_right_wrist/frame_000060.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_3/cam_high/frame_000270.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_3/cam_left_wrist/frame_000270.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_3/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the inner curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the inner curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 63.6, "hop_absolute_value": 63.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_outer_curtains_open_on_both_sides_episode_24-train-bf_000105-af_000465-plus-0000511", "task": "Pull the outer curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_20/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_20/cam_high/frame_000316.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_24/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_24/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_24/cam_right_wrist/frame_000105.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_24/cam_high/frame_000465.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_24/cam_left_wrist/frame_000465.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_24/cam_right_wrist/frame_000465.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the outer curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the outer curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_garbage_in_the_trash_can_episode_7-train-bf_000000-af_000075-plus-0000774", "task": "Dispose of garbage in the trash can", "image": [ "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_7/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_7/cam_high/frame_000693.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_7/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_7/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_7/cam_right_wrist/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_7/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_7/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_7/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of garbage in the trash can`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of garbage in the trash can`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.9, "hop_absolute_value": 10.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_39-train-bf_000705-af_000720-plus-0000757", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_28/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_28/cam_high/frame_000959.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_high/frame_000705.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_left_wrist/frame_000705.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_right_wrist/frame_000705.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_high/frame_000720.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_left_wrist/frame_000720.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_right_wrist/frame_000720.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.6, "hop_absolute_value": 2.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_sweatshirts_episode_4-train-bf_000090-af_000120-plus-0000634", "task": "Unfold the sweatshirts", "image": [ "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_10/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_10/cam_high/frame_001480.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_4/cam_high/frame_000090.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_4/cam_left_wrist/frame_000090.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_4/cam_right_wrist/frame_000090.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_4/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_4/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_4/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the sweatshirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the sweatshirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.2, "hop_absolute_value": 2.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_garbage_in_the_trash_can_episode_18-train-bf_000255-af_000270-plus-0000657", "task": "Dispose of garbage in the trash can", "image": [ "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_4/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_4/cam_high/frame_001709.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_18/cam_high/frame_000255.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_18/cam_left_wrist/frame_000255.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_18/cam_right_wrist/frame_000255.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_18/cam_high/frame_000270.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_18/cam_left_wrist/frame_000270.jpg", "images/galaxea_r1lite/Dispose_of_garbage_in_the_trash_can/episode_18/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of garbage in the trash can`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of garbage in the trash can`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.8, "hop_absolute_value": 4.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_26-train-bf_001230-af_001245-plus-0000891", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_48/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_48/cam_high/frame_001450.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_26/cam_high/frame_001230.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_26/cam_left_wrist/frame_001230.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_26/cam_right_wrist/frame_001230.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_26/cam_high/frame_001245.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_26/cam_left_wrist/frame_001245.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_26/cam_right_wrist/frame_001245.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.9, "hop_absolute_value": 1.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_shorts_episode_28-train-bf_000705-af_000750-plus-0000645", "task": "Unfold the shorts", "image": [ "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_high/frame_001883.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_high/frame_000705.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_left_wrist/frame_000705.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_right_wrist/frame_000705.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_high/frame_000750.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_left_wrist/frame_000750.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_right_wrist/frame_000750.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.8, "hop_absolute_value": 3.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_sweatshirts_episode_2-train-bf_000315-af_000330-plus-0000749", "task": "Unfold the sweatshirts", "image": [ "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_4/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_4/cam_high/frame_001477.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_2/cam_high/frame_000315.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_2/cam_left_wrist/frame_000315.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_2/cam_right_wrist/frame_000315.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_2/cam_high/frame_000330.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_2/cam_left_wrist/frame_000330.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_2/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the sweatshirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the sweatshirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.2, "hop_absolute_value": 1.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_59-train-bf_000375-af_000390-plus-0000636", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_60/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_60/cam_high/frame_001104.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_59/cam_high/frame_000375.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_59/cam_left_wrist/frame_000375.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_59/cam_right_wrist/frame_000375.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_59/cam_high/frame_000390.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_59/cam_left_wrist/frame_000390.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_59/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.9, "hop_absolute_value": 2.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_inner_curtains_on_both_sides_episode_16-train-bf_000420-af_000486-plus-0000506", "task": "Close the inner curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_12/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_12/cam_high/frame_000310.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_16/cam_high/frame_000420.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_16/cam_left_wrist/frame_000420.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_16/cam_right_wrist/frame_000420.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_16/cam_high/frame_000486.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_16/cam_left_wrist/frame_000486.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_16/cam_right_wrist/frame_000486.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the inner curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the inner curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_inner_curtains_on_both_sides_episode_18-train-bf_000060-af_000075-plus-0000285", "task": "Close the inner curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_5/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_5/cam_high/frame_000355.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_18/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_18/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_18/cam_right_wrist/frame_000060.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_18/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_18/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_18/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the inner curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the inner curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.3, "hop_absolute_value": 3.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_20-train-bf_000045-af_000075-plus-0000623", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_64/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_64/cam_high/frame_002355.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_20/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_20/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_20/cam_right_wrist/frame_000045.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_20/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_20/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_20/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.6, "hop_absolute_value": 2.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_inner_curtains_open_on_both_sides_episode_21-train-bf_000105-af_000270-plus-0000822", "task": "Pull the inner curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_high/frame_000405.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_21/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_21/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_21/cam_right_wrist/frame_000105.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_21/cam_high/frame_000270.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_21/cam_left_wrist/frame_000270.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_21/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the inner curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the inner curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 45.8, "hop_absolute_value": 45.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_sweatshirts_episode_11-train-bf_000390-af_000405-plus-0000617", "task": "Unfold the sweatshirts", "image": [ "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_6/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_6/cam_high/frame_001636.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_11/cam_high/frame_000390.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_11/cam_left_wrist/frame_000390.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_11/cam_right_wrist/frame_000390.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_11/cam_high/frame_000405.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_11/cam_left_wrist/frame_000405.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_11/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the sweatshirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the sweatshirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.4, "hop_absolute_value": 2.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_shoe_box_episode_10-train-bf_000060-af_000075-plus-0000069", "task": "Close the shoe box", "image": [ "images/galaxea_r1lite/Close_the_shoe_box/episode_10/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_10/cam_high/frame_000176.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_10/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_10/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_10/cam_right_wrist/frame_000060.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_10/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_10/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_10/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the shoe box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the shoe box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 14.3, "hop_absolute_value": 14.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_44-train-bf_001950-af_001410-minus-0000281", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_19/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_19/cam_high/frame_001221.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_44/cam_high/frame_001950.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_44/cam_left_wrist/frame_001950.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_44/cam_right_wrist/frame_001950.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_44/cam_high/frame_001410.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_44/cam_left_wrist/frame_001410.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_44/cam_right_wrist/frame_001410.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -27.7, "hop_absolute_value": 27.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Take_the_ceramic_cup_out_of_the_freezer_episode_13-train-bf_000150-af_000045-minus-0000058", "task": "Take the ceramic cup out of the freezer", "image": [ "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_13/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_13/cam_high/frame_000291.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_13/cam_high/frame_000150.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_13/cam_left_wrist/frame_000150.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_13/cam_right_wrist/frame_000150.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_13/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_13/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_13/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the ceramic cup out of the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the ceramic cup out of the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_yoghurt_in_the_freezer_episode_22-train-bf_000315-af_000300-minus-0000424", "task": "Put the yoghurt in the freezer", "image": [ "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_3/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_3/cam_high/frame_000351.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_22/cam_high/frame_000315.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_22/cam_left_wrist/frame_000315.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_22/cam_right_wrist/frame_000315.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_22/cam_high/frame_000300.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_22/cam_left_wrist/frame_000300.jpg", "images/galaxea_r1lite/Put_the_yoghurt_in_the_freezer/episode_22/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the yoghurt in the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the yoghurt in the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.8, "hop_absolute_value": 4.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_microwave_oven_episode_16-train-bf_000148-af_000000-minus-0000008", "task": "Close the microwave oven", "image": [ "images/galaxea_r1lite/Close_the_microwave_oven/episode_1/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_microwave_oven/episode_1/cam_high/frame_000188.jpg", "images/galaxea_r1lite/Close_the_microwave_oven/episode_16/cam_high/frame_000148.jpg", "images/galaxea_r1lite/Close_the_microwave_oven/episode_16/cam_left_wrist/frame_000148.jpg", "images/galaxea_r1lite/Close_the_microwave_oven/episode_16/cam_right_wrist/frame_000148.jpg", "images/galaxea_r1lite/Close_the_microwave_oven/episode_16/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_microwave_oven/episode_16/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Close_the_microwave_oven/episode_16/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the microwave oven`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the microwave oven`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_blue_chair_out_from_under_the_dining_table_episode_9-train-bf_000045-af_000030-minus-0000604", "task": "Pull the blue chair out from under the dining table", "image": [ "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_16/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_16/cam_high/frame_000372.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_9/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_9/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_9/cam_right_wrist/frame_000045.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_9/cam_high/frame_000030.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_9/cam_left_wrist/frame_000030.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_9/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the blue chair out from under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the blue chair out from under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_12-train-bf_000660-af_000615-minus-0000709", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_5/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_5/cam_high/frame_001575.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_12/cam_high/frame_000660.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_12/cam_left_wrist/frame_000660.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_12/cam_right_wrist/frame_000660.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_12/cam_high/frame_000615.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_12/cam_left_wrist/frame_000615.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_12/cam_right_wrist/frame_000615.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.8, "hop_absolute_value": 6.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_slide_door_episode_16-train-bf_000390-af_000225-minus-0000232", "task": "Open the slide door", "image": [ "images/galaxea_r1lite/Open_the_slide_door/episode_1/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_1/cam_high/frame_000669.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_16/cam_high/frame_000390.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_16/cam_left_wrist/frame_000390.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_16/cam_right_wrist/frame_000390.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_16/cam_high/frame_000225.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_16/cam_left_wrist/frame_000225.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_16/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -42.3, "hop_absolute_value": 42.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_shoe_box_episode_0-train-bf_000120-af_000060-minus-0000054", "task": "Open the shoe box", "image": [ "images/galaxea_r1lite/Open_the_shoe_box/episode_21/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_21/cam_high/frame_000333.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_0/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_0/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_0/cam_right_wrist/frame_000120.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_0/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_0/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Open_the_shoe_box/episode_0/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the shoe box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the shoe box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_outer_curtains_open_on_both_sides_episode_23-train-bf_000210-af_000060-minus-0000113", "task": "Pull the outer curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_20/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_20/cam_high/frame_000316.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_high/frame_000210.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_left_wrist/frame_000210.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_right_wrist/frame_000210.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the outer curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the outer curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_nightstand_drawer_episode_21-train-bf_000295-af_000255-minus-0000164", "task": "Open the nightstand drawer", "image": [ "images/galaxea_r1lite/Open_the_nightstand_drawer/episode_23/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_nightstand_drawer/episode_23/cam_high/frame_000169.jpg", "images/galaxea_r1lite/Open_the_nightstand_drawer/episode_21/cam_high/frame_000295.jpg", "images/galaxea_r1lite/Open_the_nightstand_drawer/episode_21/cam_left_wrist/frame_000295.jpg", "images/galaxea_r1lite/Open_the_nightstand_drawer/episode_21/cam_right_wrist/frame_000295.jpg", "images/galaxea_r1lite/Open_the_nightstand_drawer/episode_21/cam_high/frame_000255.jpg", "images/galaxea_r1lite/Open_the_nightstand_drawer/episode_21/cam_left_wrist/frame_000255.jpg", "images/galaxea_r1lite/Open_the_nightstand_drawer/episode_21/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the nightstand drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the nightstand drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.5, "hop_absolute_value": 10.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_16-train-bf_000660-af_000615-minus-0000889", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_16/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_16/cam_high/frame_001024.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_16/cam_high/frame_000660.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_16/cam_left_wrist/frame_000660.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_16/cam_right_wrist/frame_000660.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_16/cam_high/frame_000615.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_16/cam_left_wrist/frame_000615.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_16/cam_right_wrist/frame_000615.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.8, "hop_absolute_value": 6.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_leftover_food_episode_11-train-bf_001095-af_001020-minus-0000909", "task": "Dispose of leftover food", "image": [ "images/galaxea_r1lite/Dispose_of_leftover_food/episode_10/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_10/cam_high/frame_001323.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_11/cam_high/frame_001095.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_11/cam_left_wrist/frame_001095.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_11/cam_right_wrist/frame_001095.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_11/cam_high/frame_001020.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_11/cam_left_wrist/frame_001020.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_11/cam_right_wrist/frame_001020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of leftover food`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of leftover food`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.8, "hop_absolute_value": 6.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_58-train-bf_001020-af_000975-minus-0000294", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_22/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_22/cam_high/frame_000806.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_58/cam_high/frame_001020.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_58/cam_left_wrist/frame_001020.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_58/cam_right_wrist/frame_001020.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_58/cam_high/frame_000975.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_58/cam_left_wrist/frame_000975.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_58/cam_right_wrist/frame_000975.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.4, "hop_absolute_value": 4.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_inner_curtains_open_on_both_sides_episode_5-train-bf_000375-af_000135-minus-0000139", "task": "Pull the inner curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_high/frame_000432.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_high/frame_000375.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_left_wrist/frame_000375.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_right_wrist/frame_000375.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_high/frame_000135.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_left_wrist/frame_000135.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_5/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the inner curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the inner curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -64.0, "hop_absolute_value": 64.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_gray_chair_out_from_under_the_dining_table_episode_22-train-bf_000330-af_000315-minus-0000249", "task": "Pull the gray chair out from under the dining table", "image": [ "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_3/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_3/cam_high/frame_000340.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_22/cam_high/frame_000330.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_22/cam_left_wrist/frame_000330.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_22/cam_right_wrist/frame_000330.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_22/cam_high/frame_000315.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_22/cam_left_wrist/frame_000315.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_22/cam_right_wrist/frame_000315.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the gray chair out from under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the gray chair out from under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.5, "hop_absolute_value": 4.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_inner_curtains_on_both_sides_episode_14-train-bf_000240-af_000030-minus-0000041", "task": "Close the inner curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_5/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_5/cam_high/frame_000355.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_high/frame_000240.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_left_wrist/frame_000240.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_right_wrist/frame_000240.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_high/frame_000030.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_left_wrist/frame_000030.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the inner curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the inner curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_inner_curtains_open_on_both_sides_episode_24-train-bf_000045-af_000015-minus-0000052", "task": "Pull the inner curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_9/cam_high/frame_000557.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_24/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_24/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_24/cam_right_wrist/frame_000045.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_24/cam_high/frame_000015.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_24/cam_left_wrist/frame_000015.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_24/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the inner curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the inner curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_inner_curtains_open_on_both_sides_episode_18-train-bf_000195-af_000120-minus-0000125", "task": "Pull the inner curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_7/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_7/cam_high/frame_000400.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_18/cam_high/frame_000195.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_18/cam_left_wrist/frame_000195.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_18/cam_right_wrist/frame_000195.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_18/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_18/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_18/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the inner curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the inner curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -38.5, "hop_absolute_value": 38.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Push_the_blue_chair_back_under_the_dining_table_episode_10-train-bf_000390-af_000015-minus-0000029", "task": "Push the blue chair back under the dining table", "image": [ "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_21/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_21/cam_high/frame_000360.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_high/frame_000390.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_left_wrist/frame_000390.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_right_wrist/frame_000390.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_high/frame_000015.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_left_wrist/frame_000015.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the blue chair back under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the blue chair back under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -96.2, "hop_absolute_value": 96.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_inner_curtains_open_on_both_sides_episode_4-train-bf_000195-af_000075-minus-0000111", "task": "Pull the inner curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_12/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_12/cam_high/frame_000262.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_4/cam_high/frame_000195.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_4/cam_left_wrist/frame_000195.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_4/cam_right_wrist/frame_000195.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_4/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_4/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_4/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the inner curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the inner curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -61.5, "hop_absolute_value": 61.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_shorts_episode_21-train-bf_001050-af_000960-minus-0000859", "task": "Fold the shorts", "image": [ "images/galaxea_r1lite/Fold_the_shorts/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_9/cam_high/frame_000544.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_21/cam_high/frame_001050.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_21/cam_left_wrist/frame_001050.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_21/cam_right_wrist/frame_001050.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_21/cam_high/frame_000960.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_21/cam_left_wrist/frame_000960.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_21/cam_right_wrist/frame_000960.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -8.6, "hop_absolute_value": 8.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_shoes_into_the_shoe_box_episode_18-train-bf_000090-af_000075-minus-0000025", "task": "Put the shoes into the shoe box", "image": [ "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_9/cam_high/frame_000231.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_18/cam_high/frame_000090.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_18/cam_left_wrist/frame_000090.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_18/cam_right_wrist/frame_000090.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_18/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_18/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Put_the_shoes_into_the_shoe_box/episode_18/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the shoes into the shoe box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the shoes into the shoe box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_slide_door_episode_14-train-bf_000585-af_000240-minus-0000187", "task": "Open the slide door", "image": [ "images/galaxea_r1lite/Open_the_slide_door/episode_23/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_23/cam_high/frame_000607.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_14/cam_high/frame_000585.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_14/cam_left_wrist/frame_000585.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_14/cam_right_wrist/frame_000585.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_14/cam_high/frame_000240.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_14/cam_left_wrist/frame_000240.jpg", "images/galaxea_r1lite/Open_the_slide_door/episode_14/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -59.0, "hop_absolute_value": 59.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_paper_cup_in_the_freezer_episode_1-train-bf_000285-af_000075-minus-0000080", "task": "Put the paper cup in the freezer", "image": [ "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_14/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_14/cam_high/frame_000365.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_1/cam_high/frame_000285.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_1/cam_left_wrist/frame_000285.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_1/cam_right_wrist/frame_000285.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_1/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_1/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Put_the_paper_cup_in_the_freezer/episode_1/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the paper cup in the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the paper cup in the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.7, "hop_absolute_value": 73.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_sweatshirts_episode_7-train-bf_001110-af_000300-minus-0000128", "task": "Fold the sweatshirts", "image": [ "images/galaxea_r1lite/Fold_the_sweatshirts/episode_7/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_7/cam_high/frame_001856.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_7/cam_high/frame_001110.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_7/cam_left_wrist/frame_001110.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_7/cam_right_wrist/frame_001110.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_7/cam_high/frame_000300.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_7/cam_left_wrist/frame_000300.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_7/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the sweatshirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the sweatshirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.0, "hop_absolute_value": 73.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Push_the_blue_chair_back_under_the_dining_table_episode_4-train-bf_000479-af_000075-minus-0000094", "task": "Push the blue chair back under the dining table", "image": [ "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_16/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_16/cam_high/frame_000377.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_4/cam_high/frame_000479.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_4/cam_left_wrist/frame_000479.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_4/cam_right_wrist/frame_000479.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_4/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_4/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_4/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the blue chair back under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the blue chair back under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.9, "hop_absolute_value": 83.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_shorts_episode_3-train-bf_001220-af_000300-minus-0000143", "task": "Unfold the shorts", "image": [ "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_high/frame_001220.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_high/frame_001220.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_left_wrist/frame_001220.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_right_wrist/frame_001220.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_high/frame_000300.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_left_wrist/frame_000300.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.3, "hop_absolute_value": 75.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Push_the_blue_chair_back_under_the_dining_table_episode_10-train-bf_000405-af_000045-minus-0000064", "task": "Push the blue chair back under the dining table", "image": [ "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_high/frame_000439.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_high/frame_000405.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_left_wrist/frame_000405.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_right_wrist/frame_000405.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_10/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the blue chair back under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the blue chair back under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -88.9, "hop_absolute_value": 88.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_outer_curtains_on_both_sides_episode_7-train-bf_000555-af_000345-minus-0000230", "task": "Close the outer curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_23/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_23/cam_high/frame_000396.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_7/cam_high/frame_000555.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_7/cam_left_wrist/frame_000555.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_7/cam_right_wrist/frame_000555.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_7/cam_high/frame_000345.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_7/cam_left_wrist/frame_000345.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_7/cam_right_wrist/frame_000345.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the outer curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the outer curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -37.8, "hop_absolute_value": 37.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_37-train-bf_000195-af_000015-minus-0000004", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_50/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_50/cam_high/frame_001497.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_37/cam_high/frame_000195.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_37/cam_left_wrist/frame_000195.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_37/cam_right_wrist/frame_000195.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_37/cam_high/frame_000015.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_37/cam_left_wrist/frame_000015.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_37/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -92.3, "hop_absolute_value": 92.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_shorts_episode_28-train-bf_001035-af_000000-minus-0000027", "task": "Unfold the shorts", "image": [ "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_high/frame_001883.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_high/frame_001035.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_left_wrist/frame_001035.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_right_wrist/frame_001035.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_28/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_freezer_door_episode_9-train-bf_000135-af_000120-minus-0000130", "task": "Close the freezer door", "image": [ "images/galaxea_r1lite/Close_the_freezer_door/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_freezer_door/episode_9/cam_high/frame_000252.jpg", "images/galaxea_r1lite/Close_the_freezer_door/episode_9/cam_high/frame_000135.jpg", "images/galaxea_r1lite/Close_the_freezer_door/episode_9/cam_left_wrist/frame_000135.jpg", "images/galaxea_r1lite/Close_the_freezer_door/episode_9/cam_right_wrist/frame_000135.jpg", "images/galaxea_r1lite/Close_the_freezer_door/episode_9/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Close_the_freezer_door/episode_9/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Close_the_freezer_door/episode_9/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the freezer door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the freezer door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_blue_chair_out_from_under_the_dining_table_episode_21-train-bf_000135-af_000090-minus-0000154", "task": "Pull the blue chair out from under the dining table", "image": [ "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_19/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_19/cam_high/frame_000482.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_21/cam_high/frame_000135.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_21/cam_left_wrist/frame_000135.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_21/cam_right_wrist/frame_000135.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_21/cam_high/frame_000090.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_21/cam_left_wrist/frame_000090.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_21/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the blue chair out from under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the blue chair out from under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_24-train-bf_000225-af_000210-minus-0000618", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_24/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_24/cam_high/frame_000588.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_24/cam_high/frame_000225.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_24/cam_left_wrist/frame_000225.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_24/cam_right_wrist/frame_000225.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_24/cam_high/frame_000210.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_24/cam_left_wrist/frame_000210.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_24/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.7, "hop_absolute_value": 6.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Take_the_data_cable_out_of_the_drawer_and_put_it_on_the_table_episode_5-train-bf_000120-af_000075-minus-0000044", "task": "Take the data cable out of the drawer and put it on the table", "image": [ "images/galaxea_r1lite/Take_the_data_cable_out_of_the_drawer_and_put_it_on_the_table/episode_5/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_data_cable_out_of_the_drawer_and_put_it_on_the_table/episode_5/cam_high/frame_000174.jpg", "images/galaxea_r1lite/Take_the_data_cable_out_of_the_drawer_and_put_it_on_the_table/episode_5/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Take_the_data_cable_out_of_the_drawer_and_put_it_on_the_table/episode_5/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Take_the_data_cable_out_of_the_drawer_and_put_it_on_the_table/episode_5/cam_right_wrist/frame_000120.jpg", "images/galaxea_r1lite/Take_the_data_cable_out_of_the_drawer_and_put_it_on_the_table/episode_5/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Take_the_data_cable_out_of_the_drawer_and_put_it_on_the_table/episode_5/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Take_the_data_cable_out_of_the_drawer_and_put_it_on_the_table/episode_5/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the data cable out of the drawer and put it on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the data cable out of the drawer and put it on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -37.5, "hop_absolute_value": 37.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_leftover_food_episode_20-train-bf_000420-af_000375-minus-0000715", "task": "Dispose of leftover food", "image": [ "images/galaxea_r1lite/Dispose_of_leftover_food/episode_23/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_23/cam_high/frame_000876.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_20/cam_high/frame_000420.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_20/cam_left_wrist/frame_000420.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_20/cam_right_wrist/frame_000420.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_20/cam_high/frame_000375.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_20/cam_left_wrist/frame_000375.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_20/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of leftover food`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of leftover food`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.7, "hop_absolute_value": 10.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Take_the_yoghurt_out_of_the_freezer_episode_10-train-bf_000135-af_000060-minus-0000028", "task": "Take the yoghurt out of the freezer", "image": [ "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_24/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_24/cam_high/frame_000175.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_10/cam_high/frame_000135.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_10/cam_left_wrist/frame_000135.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_10/cam_right_wrist/frame_000135.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_10/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_10/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Take_the_yoghurt_out_of_the_freezer/episode_10/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the yoghurt out of the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the yoghurt out of the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -55.6, "hop_absolute_value": 55.6, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_12-train-bf_000720-af_000645-minus-0000730", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_1/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_1/cam_high/frame_001472.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_12/cam_high/frame_000720.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_12/cam_left_wrist/frame_000720.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_12/cam_right_wrist/frame_000720.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_12/cam_high/frame_000645.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_12/cam_left_wrist/frame_000645.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_12/cam_right_wrist/frame_000645.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.4, "hop_absolute_value": 10.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Take_the_ceramic_cup_out_of_the_freezer_episode_10-train-bf_000270-af_000165-minus-0000126", "task": "Take the ceramic cup out of the freezer", "image": [ "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_10/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_10/cam_high/frame_000290.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_10/cam_high/frame_000270.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_10/cam_left_wrist/frame_000270.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_10/cam_right_wrist/frame_000270.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_10/cam_high/frame_000165.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_10/cam_left_wrist/frame_000165.jpg", "images/galaxea_r1lite/Take_the_ceramic_cup_out_of_the_freezer/episode_10/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the ceramic cup out of the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the ceramic cup out of the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -38.9, "hop_absolute_value": 38.9, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_46-train-bf_001605-af_000090-minus-0000029", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_50/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_50/cam_high/frame_001497.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_46/cam_high/frame_001605.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_46/cam_left_wrist/frame_001605.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_46/cam_right_wrist/frame_001605.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_46/cam_high/frame_000090.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_46/cam_left_wrist/frame_000090.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_46/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -94.4, "hop_absolute_value": 94.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_shoe_box_episode_20-train-bf_000255-af_000210-minus-0000876", "task": "Close the shoe box", "image": [ "images/galaxea_r1lite/Close_the_shoe_box/episode_11/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_11/cam_high/frame_000274.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_20/cam_high/frame_000255.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_20/cam_left_wrist/frame_000255.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_20/cam_right_wrist/frame_000255.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_20/cam_high/frame_000210.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_20/cam_left_wrist/frame_000210.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_20/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the shoe box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the shoe box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -17.6, "hop_absolute_value": 17.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_drawer_episode_1-train-bf_000045-af_000030-minus-0000136", "task": "Open the drawer", "image": [ "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_high/frame_000287.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_right_wrist/frame_000045.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_high/frame_000030.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_left_wrist/frame_000030.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_gray_chair_out_from_under_the_dining_table_episode_17-train-bf_000360-af_000060-minus-0000086", "task": "Pull the gray chair out from under the dining table", "image": [ "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_8/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_8/cam_high/frame_000234.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_high/frame_000360.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_left_wrist/frame_000360.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_right_wrist/frame_000360.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_17/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the gray chair out from under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the gray chair out from under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_gray_chair_out_from_under_the_dining_table_episode_5-train-bf_000525-af_000000-minus-0000029", "task": "Pull the gray chair out from under the dining table", "image": [ "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_5/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_5/cam_high/frame_000646.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_5/cam_high/frame_000525.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_5/cam_left_wrist/frame_000525.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_5/cam_right_wrist/frame_000525.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_5/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_5/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_gray_chair_out_from_under_the_dining_table/episode_5/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the gray chair out from under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the gray chair out from under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_29-train-bf_000870-af_000375-minus-0000198", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_29/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_29/cam_high/frame_000965.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_29/cam_high/frame_000870.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_29/cam_left_wrist/frame_000870.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_29/cam_right_wrist/frame_000870.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_29/cam_high/frame_000375.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_29/cam_left_wrist/frame_000375.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_29/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -56.9, "hop_absolute_value": 56.9, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_39-train-bf_001275-af_001170-minus-0000297", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_41/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_41/cam_high/frame_000918.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_high/frame_001275.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_left_wrist/frame_001275.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_right_wrist/frame_001275.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_high/frame_001170.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_left_wrist/frame_001170.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_right_wrist/frame_001170.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -8.2, "hop_absolute_value": 8.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_inner_curtains_on_both_sides_episode_14-train-bf_000300-af_000075-minus-0000099", "task": "Close the inner curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_3/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_3/cam_high/frame_000415.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_high/frame_000300.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_left_wrist/frame_000300.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_right_wrist/frame_000300.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_14/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the inner curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the inner curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table_episode_1-train-bf_000480-af_000375-minus-0000751", "task": "Take the tissue out of the drawer one by one and place them on the table", "image": [ "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_high/frame_000548.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_high/frame_000480.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_left_wrist/frame_000480.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_right_wrist/frame_000480.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_high/frame_000375.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_left_wrist/frame_000375.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the tissue out of the drawer one by one and place them on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the tissue out of the drawer one by one and place them on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -21.9, "hop_absolute_value": 21.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_inner_curtains_open_on_both_sides_episode_15-train-bf_000030-af_000015-minus-0000667", "task": "Pull the inner curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_17/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_17/cam_high/frame_000441.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_high/frame_000030.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_left_wrist/frame_000030.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_right_wrist/frame_000030.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_high/frame_000015.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_left_wrist/frame_000015.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the inner curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the inner curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_shorts_episode_9-train-bf_000495-af_000390-minus-0000756", "task": "Fold the shorts", "image": [ "images/galaxea_r1lite/Fold_the_shorts/episode_7/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_7/cam_high/frame_001428.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_9/cam_high/frame_000495.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_9/cam_left_wrist/frame_000495.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_9/cam_right_wrist/frame_000495.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_9/cam_high/frame_000390.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_9/cam_left_wrist/frame_000390.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_9/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -21.2, "hop_absolute_value": 21.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_shorts_episode_28-train-bf_000360-af_000240-minus-0000829", "task": "Fold the shorts", "image": [ "images/galaxea_r1lite/Fold_the_shorts/episode_14/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_14/cam_high/frame_000614.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_28/cam_high/frame_000360.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_28/cam_left_wrist/frame_000360.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_28/cam_right_wrist/frame_000360.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_28/cam_high/frame_000240.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_28/cam_left_wrist/frame_000240.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_28/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table_episode_5-train-bf_000405-af_000135-minus-0000139", "task": "Take the remote control out of the drawer one by one and place them on the table", "image": [ "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_4/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_4/cam_high/frame_000547.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_5/cam_high/frame_000405.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_5/cam_left_wrist/frame_000405.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_5/cam_right_wrist/frame_000405.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_5/cam_high/frame_000135.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_5/cam_left_wrist/frame_000135.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_5/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the remote control out of the drawer one by one and place them on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the remote control out of the drawer one by one and place them on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_outer_curtains_on_both_sides_episode_6-train-bf_000375-af_000255-minus-0000883", "task": "Close the outer curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_8/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_8/cam_high/frame_000425.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_6/cam_high/frame_000375.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_6/cam_left_wrist/frame_000375.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_6/cam_right_wrist/frame_000375.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_6/cam_high/frame_000255.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_6/cam_left_wrist/frame_000255.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_6/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the outer curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the outer curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -32.0, "hop_absolute_value": 32.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_52-train-bf_000975-af_000945-minus-0000746", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_38/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_38/cam_high/frame_001197.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_52/cam_high/frame_000975.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_52/cam_left_wrist/frame_000975.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_52/cam_right_wrist/frame_000975.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_52/cam_high/frame_000945.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_52/cam_left_wrist/frame_000945.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_52/cam_right_wrist/frame_000945.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.1, "hop_absolute_value": 3.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_sweatshirts_episode_4-train-bf_000630-af_000390-minus-0000237", "task": "Fold the sweatshirts", "image": [ "images/galaxea_r1lite/Fold_the_sweatshirts/episode_2/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_2/cam_high/frame_001811.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_4/cam_high/frame_000630.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_4/cam_left_wrist/frame_000630.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_4/cam_right_wrist/frame_000630.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_4/cam_high/frame_000390.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_4/cam_left_wrist/frame_000390.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_4/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the sweatshirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the sweatshirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -38.1, "hop_absolute_value": 38.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_10-train-bf_000690-af_000000-minus-0000027", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_40/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_40/cam_high/frame_000918.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_10/cam_high/frame_000690.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_10/cam_left_wrist/frame_000690.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_10/cam_right_wrist/frame_000690.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_10/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_10/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_10/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_leftover_food_episode_3-train-bf_001080-af_000405-minus-0000164", "task": "Dispose of leftover food", "image": [ "images/galaxea_r1lite/Dispose_of_leftover_food/episode_10/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_10/cam_high/frame_001323.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_3/cam_high/frame_001080.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_3/cam_left_wrist/frame_001080.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_3/cam_right_wrist/frame_001080.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_3/cam_high/frame_000405.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_3/cam_left_wrist/frame_000405.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_3/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of leftover food`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of leftover food`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_44-train-bf_000495-af_000450-minus-0000746", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_31/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_31/cam_high/frame_000959.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_44/cam_high/frame_000495.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_44/cam_left_wrist/frame_000495.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_44/cam_right_wrist/frame_000495.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_44/cam_high/frame_000450.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_44/cam_left_wrist/frame_000450.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_44/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_inner_curtains_open_on_both_sides_episode_15-train-bf_000405-af_000180-minus-0000154", "task": "Pull the inner curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_3/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_3/cam_high/frame_000399.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_high/frame_000405.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_left_wrist/frame_000405.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_right_wrist/frame_000405.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_high/frame_000180.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_left_wrist/frame_000180.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_15/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the inner curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the inner curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -55.6, "hop_absolute_value": 55.6, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table_episode_1-train-bf_000548-af_000180-minus-0000153", "task": "Take the tissue out of the drawer one by one and place them on the table", "image": [ "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_4/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_4/cam_high/frame_000174.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_high/frame_000548.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_left_wrist/frame_000548.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_right_wrist/frame_000548.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_high/frame_000180.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_left_wrist/frame_000180.jpg", "images/galaxea_r1lite/Take_the_tissue_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_1/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the tissue out of the drawer one by one and place them on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the tissue out of the drawer one by one and place them on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table_episode_5-train-bf_000225-af_000030-minus-0000039", "task": "Take the remote control out of the drawer one by one and place them on the table", "image": [ "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_22/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_22/cam_high/frame_000325.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_5/cam_high/frame_000225.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_5/cam_left_wrist/frame_000225.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_5/cam_right_wrist/frame_000225.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_5/cam_high/frame_000030.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_5/cam_left_wrist/frame_000030.jpg", "images/galaxea_r1lite/Take_the_remote_control_out_of_the_drawer_one_by_one_and_place_them_on_the_table/episode_5/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the remote control out of the drawer one by one and place them on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the remote control out of the drawer one by one and place them on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -86.7, "hop_absolute_value": 86.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_microwave_oven_episode_11-train-bf_000135-af_000105-minus-0000199", "task": "Open the microwave oven", "image": [ "images/galaxea_r1lite/Open_the_microwave_oven/episode_18/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_18/cam_high/frame_000209.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_11/cam_high/frame_000135.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_11/cam_left_wrist/frame_000135.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_11/cam_right_wrist/frame_000135.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_11/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_11/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_11/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the microwave oven`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the microwave oven`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -22.2, "hop_absolute_value": 22.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one_episode_6-train-bf_000315-af_000270-minus-0000586", "task": "Put the remote controls on the table in the drawer one by one", "image": [ "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_17/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_17/cam_high/frame_000414.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_6/cam_high/frame_000315.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_6/cam_left_wrist/frame_000315.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_6/cam_right_wrist/frame_000315.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_6/cam_high/frame_000270.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_6/cam_left_wrist/frame_000270.jpg", "images/galaxea_r1lite/Put_the_remote_controls_on_the_table_in_the_drawer_one_by_one/episode_6/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the remote controls on the table in the drawer one by one`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the remote controls on the table in the drawer one by one`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -14.3, "hop_absolute_value": 14.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_slide_door_episode_14-train-bf_000630-af_000570-minus-0000734", "task": "Close the slide door", "image": [ "images/galaxea_r1lite/Close_the_slide_door/episode_21/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_21/cam_high/frame_000352.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_14/cam_high/frame_000630.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_14/cam_left_wrist/frame_000630.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_14/cam_right_wrist/frame_000630.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_14/cam_high/frame_000570.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_14/cam_left_wrist/frame_000570.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_14/cam_right_wrist/frame_000570.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -9.5, "hop_absolute_value": 9.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_plate_in_the_microwave_oven_episode_24-train-bf_000105-af_000075-minus-0000111", "task": "Put the plate in the microwave oven", "image": [ "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_24/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_24/cam_high/frame_000269.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_24/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_24/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_24/cam_right_wrist/frame_000105.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_24/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_24/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Put_the_plate_in_the_microwave_oven/episode_24/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the plate in the microwave oven`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the plate in the microwave oven`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -28.6, "hop_absolute_value": 28.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_sweatshirts_episode_7-train-bf_000510-af_000255-minus-0000214", "task": "Unfold the sweatshirts", "image": [ "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_5/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_5/cam_high/frame_001556.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_7/cam_high/frame_000510.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_7/cam_left_wrist/frame_000510.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_7/cam_right_wrist/frame_000510.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_7/cam_high/frame_000255.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_7/cam_left_wrist/frame_000255.jpg", "images/galaxea_r1lite/Unfold_the_sweatshirts/episode_7/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the sweatshirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the sweatshirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Place_the_paper_cup_on_the_table_episode_4-train-bf_000120-af_000060-minus-0000033", "task": "Place the paper cup on the table", "image": [ "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_9/cam_high/frame_000178.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_4/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_4/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_4/cam_right_wrist/frame_000120.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_4/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_4/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_4/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the paper cup on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the paper cup on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Push_the_blue_chair_back_under_the_dining_table_episode_11-train-bf_000075-af_000015-minus-0000047", "task": "Push the blue chair back under the dining table", "image": [ "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_16/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_16/cam_high/frame_000377.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_11/cam_high/frame_000075.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_11/cam_left_wrist/frame_000075.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_11/cam_right_wrist/frame_000075.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_11/cam_high/frame_000015.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_11/cam_left_wrist/frame_000015.jpg", "images/galaxea_r1lite/Push_the_blue_chair_back_under_the_dining_table/episode_11/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Push the blue chair back under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Push the blue chair back under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_drawer_episode_1-train-bf_000180-af_000045-minus-0000059", "task": "Open the drawer", "image": [ "images/galaxea_r1lite/Open_the_drawer/episode_14/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_14/cam_high/frame_000271.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_high/frame_000180.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_left_wrist/frame_000180.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_right_wrist/frame_000180.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Open_the_drawer/episode_1/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_inner_curtains_open_on_both_sides_episode_7-train-bf_000165-af_000060-minus-0000115", "task": "Pull the inner curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_9/cam_high/frame_000557.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_7/cam_high/frame_000165.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_7/cam_left_wrist/frame_000165.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_7/cam_right_wrist/frame_000165.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_7/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_7/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Pull_the_inner_curtains_open_on_both_sides/episode_7/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the inner curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the inner curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -63.6, "hop_absolute_value": 63.6, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_inner_curtains_on_both_sides_episode_1-train-bf_000420-af_000150-minus-0000150", "task": "Close the inner curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_9/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_9/cam_high/frame_000466.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_1/cam_high/frame_000420.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_1/cam_left_wrist/frame_000420.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_1/cam_right_wrist/frame_000420.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_1/cam_high/frame_000150.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_1/cam_left_wrist/frame_000150.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_1/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the inner curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the inner curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -64.3, "hop_absolute_value": 64.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Place_the_ceramic_cup_on_the_table_episode_13-train-bf_000060-af_000015-minus-0000032", "task": "Place the ceramic cup on the table", "image": [ "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_8/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_8/cam_high/frame_000214.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_13/cam_high/frame_000060.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_13/cam_left_wrist/frame_000060.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_13/cam_right_wrist/frame_000060.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_13/cam_high/frame_000015.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_13/cam_left_wrist/frame_000015.jpg", "images/galaxea_r1lite/Place_the_ceramic_cup_on_the_table/episode_13/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the ceramic cup on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the ceramic cup on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_inner_curtains_on_both_sides_episode_20-train-bf_000330-af_000225-minus-0000633", "task": "Close the inner curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_11/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_11/cam_high/frame_000519.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_20/cam_high/frame_000330.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_20/cam_left_wrist/frame_000330.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_20/cam_right_wrist/frame_000330.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_20/cam_high/frame_000225.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_20/cam_left_wrist/frame_000225.jpg", "images/galaxea_r1lite/Close_the_inner_curtains_on_both_sides/episode_20/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the inner curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the inner curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -31.8, "hop_absolute_value": 31.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_shorts_episode_6-train-bf_000045-af_000030-minus-0000235", "task": "Fold the shorts", "image": [ "images/galaxea_r1lite/Fold_the_shorts/episode_20/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_20/cam_high/frame_000899.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_6/cam_high/frame_000045.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_6/cam_left_wrist/frame_000045.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_6/cam_right_wrist/frame_000045.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_6/cam_high/frame_000030.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_6/cam_left_wrist/frame_000030.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_6/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_63-train-bf_001185-af_001110-minus-0000939", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_37/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_37/cam_high/frame_001594.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_63/cam_high/frame_001185.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_63/cam_left_wrist/frame_001185.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_63/cam_right_wrist/frame_001185.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_63/cam_high/frame_001110.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_63/cam_left_wrist/frame_001110.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_63/cam_right_wrist/frame_001110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.3, "hop_absolute_value": 6.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Open_the_microwave_oven_episode_6-train-bf_000105-af_000000-minus-0000008", "task": "Open the microwave oven", "image": [ "images/galaxea_r1lite/Open_the_microwave_oven/episode_21/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_21/cam_high/frame_000286.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_6/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_6/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_6/cam_right_wrist/frame_000105.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_6/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_6/cam_left_wrist/frame_000000.jpg", "images/galaxea_r1lite/Open_the_microwave_oven/episode_6/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the microwave oven`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the microwave oven`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_leftover_food_episode_17-train-bf_001185-af_000105-minus-0000069", "task": "Dispose of leftover food", "image": [ "images/galaxea_r1lite/Dispose_of_leftover_food/episode_11/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_11/cam_high/frame_001299.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_17/cam_high/frame_001185.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_17/cam_left_wrist/frame_001185.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_17/cam_right_wrist/frame_001185.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_17/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_17/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_17/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of leftover food`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of leftover food`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -91.1, "hop_absolute_value": 91.1, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_outer_curtains_on_both_sides_episode_9-train-bf_000529-af_000495-minus-0000612", "task": "Close the outer curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_15/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_15/cam_high/frame_000443.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_9/cam_high/frame_000529.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_9/cam_left_wrist/frame_000529.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_9/cam_right_wrist/frame_000529.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_9/cam_high/frame_000495.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_9/cam_left_wrist/frame_000495.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_9/cam_right_wrist/frame_000495.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the outer curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the outer curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.7, "hop_absolute_value": 5.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_slide_door_episode_10-train-bf_000600-af_000585-minus-0000595", "task": "Close the slide door", "image": [ "images/galaxea_r1lite/Close_the_slide_door/episode_16/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_16/cam_high/frame_000545.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_10/cam_high/frame_000600.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_10/cam_left_wrist/frame_000600.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_10/cam_right_wrist/frame_000600.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_10/cam_high/frame_000585.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_10/cam_left_wrist/frame_000585.jpg", "images/galaxea_r1lite/Close_the_slide_door/episode_10/cam_right_wrist/frame_000585.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the slide door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the slide door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.5, "hop_absolute_value": 2.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_t-shirts_episode_50-train-bf_000735-af_000720-minus-0000867", "task": "Fold the t-shirts", "image": [ "images/galaxea_r1lite/Fold_the_t-shirts/episode_31/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_31/cam_high/frame_001773.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_50/cam_high/frame_000735.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_50/cam_left_wrist/frame_000735.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_50/cam_right_wrist/frame_000735.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_50/cam_high/frame_000720.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_50/cam_left_wrist/frame_000720.jpg", "images/galaxea_r1lite/Fold_the_t-shirts/episode_50/cam_right_wrist/frame_000720.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.0, "hop_absolute_value": 2.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_outer_curtains_on_both_sides_episode_24-train-bf_000315-af_000225-minus-0000238", "task": "Close the outer curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_15/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_15/cam_high/frame_000443.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_24/cam_high/frame_000315.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_24/cam_left_wrist/frame_000315.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_24/cam_right_wrist/frame_000315.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_24/cam_high/frame_000225.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_24/cam_left_wrist/frame_000225.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_24/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the outer curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the outer curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -28.6, "hop_absolute_value": 28.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_sweatshirts_episode_10-train-bf_001215-af_001170-minus-0000314", "task": "Fold the sweatshirts", "image": [ "images/galaxea_r1lite/Fold_the_sweatshirts/episode_11/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_11/cam_high/frame_002071.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_10/cam_high/frame_001215.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_10/cam_left_wrist/frame_001215.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_10/cam_right_wrist/frame_001215.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_10/cam_high/frame_001170.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_10/cam_left_wrist/frame_001170.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_10/cam_right_wrist/frame_001170.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the sweatshirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the sweatshirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.7, "hop_absolute_value": 3.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_sweatshirts_episode_9-train-bf_000645-af_000600-minus-0000916", "task": "Fold the sweatshirts", "image": [ "images/galaxea_r1lite/Fold_the_sweatshirts/episode_11/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_11/cam_high/frame_002071.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_9/cam_high/frame_000645.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_9/cam_left_wrist/frame_000645.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_9/cam_right_wrist/frame_000645.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_9/cam_high/frame_000600.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_9/cam_left_wrist/frame_000600.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_9/cam_right_wrist/frame_000600.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the sweatshirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the sweatshirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.0, "hop_absolute_value": 7.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_shorts_episode_11-train-bf_000810-af_000780-minus-0000661", "task": "Unfold the shorts", "image": [ "images/galaxea_r1lite/Unfold_the_shorts/episode_5/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_5/cam_high/frame_000701.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_11/cam_high/frame_000810.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_11/cam_left_wrist/frame_000810.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_11/cam_right_wrist/frame_000810.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_11/cam_high/frame_000780.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_11/cam_left_wrist/frame_000780.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_11/cam_right_wrist/frame_000780.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.7, "hop_absolute_value": 3.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_39-train-bf_000285-af_000270-minus-0000824", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_high/frame_001295.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_high/frame_000285.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_left_wrist/frame_000285.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_right_wrist/frame_000285.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_high/frame_000270.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_left_wrist/frame_000270.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_39/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.3, "hop_absolute_value": 5.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_outer_curtains_on_both_sides_episode_11-train-bf_000120-af_000105-minus-0000722", "task": "Close the outer curtains on both sides", "image": [ "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_4/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_4/cam_high/frame_000632.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_11/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_11/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_11/cam_right_wrist/frame_000120.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_11/cam_high/frame_000105.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_11/cam_left_wrist/frame_000105.jpg", "images/galaxea_r1lite/Close_the_outer_curtains_on_both_sides/episode_11/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the outer curtains on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the outer curtains on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_shorts_episode_22-train-bf_000840-af_000825-minus-0000938", "task": "Fold the shorts", "image": [ "images/galaxea_r1lite/Fold_the_shorts/episode_22/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_22/cam_high/frame_001725.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_22/cam_high/frame_000840.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_22/cam_left_wrist/frame_000840.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_22/cam_right_wrist/frame_000840.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_22/cam_high/frame_000825.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_22/cam_left_wrist/frame_000825.jpg", "images/galaxea_r1lite/Fold_the_shorts/episode_22/cam_right_wrist/frame_000825.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.8, "hop_absolute_value": 1.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_outer_curtains_open_on_both_sides_episode_23-train-bf_000330-af_000315-minus-0000609", "task": "Pull the outer curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_17/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_17/cam_high/frame_000371.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_high/frame_000330.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_left_wrist/frame_000330.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_right_wrist/frame_000330.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_high/frame_000315.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_left_wrist/frame_000315.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_23/cam_right_wrist/frame_000315.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the outer curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the outer curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.5, "hop_absolute_value": 4.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Place_the_paper_cup_on_the_table_episode_15-train-bf_000195-af_000180-minus-0000101", "task": "Place the paper cup on the table", "image": [ "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_15/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_15/cam_high/frame_000213.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_15/cam_high/frame_000195.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_15/cam_left_wrist/frame_000195.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_15/cam_right_wrist/frame_000195.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_15/cam_high/frame_000180.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_15/cam_left_wrist/frame_000180.jpg", "images/galaxea_r1lite/Place_the_paper_cup_on_the_table/episode_15/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the paper cup on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the paper cup on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.7, "hop_absolute_value": 7.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Put_the_ceramic_cup_in_the_freezer_episode_4-train-bf_000585-af_000555-minus-0000625", "task": "Put the ceramic cup in the freezer", "image": [ "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_8/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_8/cam_high/frame_000414.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_4/cam_high/frame_000585.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_4/cam_left_wrist/frame_000585.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_4/cam_right_wrist/frame_000585.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_4/cam_high/frame_000555.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_4/cam_left_wrist/frame_000555.jpg", "images/galaxea_r1lite/Put_the_ceramic_cup_in_the_freezer/episode_4/cam_right_wrist/frame_000555.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put the ceramic cup in the freezer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put the ceramic cup in the freezer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.1, "hop_absolute_value": 5.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_shorts_episode_23-train-bf_000765-af_000735-minus-0000785", "task": "Unfold the shorts", "image": [ "images/galaxea_r1lite/Unfold_the_shorts/episode_13/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_13/cam_high/frame_001646.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_23/cam_high/frame_000765.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_23/cam_left_wrist/frame_000765.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_23/cam_right_wrist/frame_000765.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_23/cam_high/frame_000735.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_23/cam_left_wrist/frame_000735.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_23/cam_right_wrist/frame_000735.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.9, "hop_absolute_value": 3.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_leftover_food_episode_5-train-bf_000495-af_000480-minus-0000645", "task": "Dispose of leftover food", "image": [ "images/galaxea_r1lite/Dispose_of_leftover_food/episode_5/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_5/cam_high/frame_000851.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_5/cam_high/frame_000495.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_5/cam_left_wrist/frame_000495.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_5/cam_right_wrist/frame_000495.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_5/cam_high/frame_000480.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_5/cam_left_wrist/frame_000480.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_5/cam_right_wrist/frame_000480.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of leftover food`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of leftover food`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.0, "hop_absolute_value": 3.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_outer_curtains_open_on_both_sides_episode_2-train-bf_000225-af_000210-minus-0000128", "task": "Pull the outer curtains open on both sides", "image": [ "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_17/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_17/cam_high/frame_000371.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_2/cam_high/frame_000225.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_2/cam_left_wrist/frame_000225.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_2/cam_right_wrist/frame_000225.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_2/cam_high/frame_000210.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_2/cam_left_wrist/frame_000210.jpg", "images/galaxea_r1lite/Pull_the_outer_curtains_open_on_both_sides/episode_2/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the outer curtains open on both sides`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the outer curtains open on both sides`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.7, "hop_absolute_value": 6.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Close_the_shoe_box_episode_7-train-bf_000135-af_000120-minus-0000492", "task": "Close the shoe box", "image": [ "images/galaxea_r1lite/Close_the_shoe_box/episode_6/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_6/cam_high/frame_000706.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_7/cam_high/frame_000135.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_7/cam_left_wrist/frame_000135.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_7/cam_right_wrist/frame_000135.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_7/cam_high/frame_000120.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_7/cam_left_wrist/frame_000120.jpg", "images/galaxea_r1lite/Close_the_shoe_box/episode_7/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Close the shoe box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Close the shoe box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_shorts_episode_3-train-bf_000405-af_000360-minus-0000877", "task": "Unfold the shorts", "image": [ "images/galaxea_r1lite/Unfold_the_shorts/episode_8/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_8/cam_high/frame_001658.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_high/frame_000405.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_left_wrist/frame_000405.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_right_wrist/frame_000405.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_high/frame_000360.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_left_wrist/frame_000360.jpg", "images/galaxea_r1lite/Unfold_the_shorts/episode_3/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the shorts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the shorts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Dispose_of_leftover_food_episode_8-train-bf_001218-af_001200-minus-0000722", "task": "Dispose of leftover food", "image": [ "images/galaxea_r1lite/Dispose_of_leftover_food/episode_18/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_18/cam_high/frame_000927.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_8/cam_high/frame_001218.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_8/cam_left_wrist/frame_001218.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_8/cam_right_wrist/frame_001218.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_8/cam_high/frame_001200.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_8/cam_left_wrist/frame_001200.jpg", "images/galaxea_r1lite/Dispose_of_leftover_food/episode_8/cam_right_wrist/frame_001200.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Dispose of leftover food`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Dispose of leftover food`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.2, "hop_absolute_value": 1.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_2-train-bf_000300-af_000285-minus-0000792", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_16/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_16/cam_high/frame_001285.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_2/cam_high/frame_000300.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_2/cam_left_wrist/frame_000300.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_2/cam_right_wrist/frame_000300.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_2/cam_high/frame_000285.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_2/cam_left_wrist/frame_000285.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_2/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.0, "hop_absolute_value": 5.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Fold_the_sweatshirts_episode_10-train-bf_000525-af_000510-minus-0000706", "task": "Fold the sweatshirts", "image": [ "images/galaxea_r1lite/Fold_the_sweatshirts/episode_13/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_13/cam_high/frame_001525.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_10/cam_high/frame_000525.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_10/cam_left_wrist/frame_000525.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_10/cam_right_wrist/frame_000525.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_10/cam_high/frame_000510.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_10/cam_left_wrist/frame_000510.jpg", "images/galaxea_r1lite/Fold_the_sweatshirts/episode_10/cam_right_wrist/frame_000510.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the sweatshirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the sweatshirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.9, "hop_absolute_value": 2.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Unfold_the_t-shirts_episode_20-train-bf_000660-af_000645-minus-0000700", "task": "Unfold the t-shirts", "image": [ "images/galaxea_r1lite/Unfold_the_t-shirts/episode_53/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_53/cam_high/frame_001552.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_20/cam_high/frame_000660.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_20/cam_left_wrist/frame_000660.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_20/cam_right_wrist/frame_000660.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_20/cam_high/frame_000645.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_20/cam_left_wrist/frame_000645.jpg", "images/galaxea_r1lite/Unfold_the_t-shirts/episode_20/cam_right_wrist/frame_000645.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the t-shirts`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the t-shirts`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.3, "hop_absolute_value": 2.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "galaxea_r1lite_Pull_the_blue_chair_out_from_under_the_dining_table_episode_24-train-bf_000240-af_000210-minus-0000546", "task": "Pull the blue chair out from under the dining table", "image": [ "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_1/cam_high/frame_000000.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_1/cam_high/frame_000360.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_24/cam_high/frame_000240.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_24/cam_left_wrist/frame_000240.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_24/cam_right_wrist/frame_000240.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_24/cam_high/frame_000210.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_24/cam_left_wrist/frame_000210.jpg", "images/galaxea_r1lite/Pull_the_blue_chair_out_from_under_the_dining_table/episode_24/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the blue chair out from under the dining table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the blue chair out from under the dining table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "galaxea_r1lite" }, { "id": "human_egodex_Unwrap_a_chickenburger_from_foil_on_a_pink_tablecloth_episode_17-train-bf_000075-af_000150-plus-0000281", "task": "Unwrap a chickenburger from foil on a pink tablecloth", "image": [ "images/human_egodex/Unwrap_a_chickenburger_from_foil_on_a_pink_tablecloth/episode_17/cam_high/frame_000000.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_foil_on_a_pink_tablecloth/episode_17/cam_high/frame_000490.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_foil_on_a_pink_tablecloth/episode_17/cam_high/frame_000075.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_foil_on_a_pink_tablecloth/episode_17/cam_high/frame_000075.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_foil_on_a_pink_tablecloth/episode_17/cam_high/frame_000075.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_foil_on_a_pink_tablecloth/episode_17/cam_high/frame_000150.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_foil_on_a_pink_tablecloth/episode_17/cam_high/frame_000150.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_foil_on_a_pink_tablecloth/episode_17/cam_high/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unwrap a chickenburger from foil on a pink tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unwrap a chickenburger from foil on a pink tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 18.8, "hop_absolute_value": 18.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Screw_the_screw_into_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting_episode_3-train-bf_000075-af_000225-plus-0000163", "task": "Screw the screw into the fixture using an allen key on a pink tablecloth with a lavender background while sitting", "image": [ "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_12/cam_high/frame_000000.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_12/cam_high/frame_000803.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_3/cam_high/frame_000075.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_3/cam_high/frame_000075.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_3/cam_high/frame_000075.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_3/cam_high/frame_000225.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_3/cam_high/frame_000225.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_3/cam_high/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Screw the screw into the fixture using an allen key on a pink tablecloth with a lavender background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Screw the screw into the fixture using an allen key on a pink tablecloth with a lavender background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 23.1, "hop_absolute_value": 23.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Open_the_box_remove_the_five_duck_objects_and_five_plush_objects_from_it_then_close_the_box_on_a_wooden_table_with_a_white_background_episode_17-train-bf_000400-af_000427-plus-0000220", "task": "Open the box remove the five duck objects and five plush objects from it then close the box on a wooden table with a white background", "image": [ "images/human_egodex/Open_the_box_remove_the_five_duck_objects_and_five_plush_objects_from_it_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_17/cam_high/frame_000000.jpg", "images/human_egodex/Open_the_box_remove_the_five_duck_objects_and_five_plush_objects_from_it_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_17/cam_high/frame_000427.jpg", "images/human_egodex/Open_the_box_remove_the_five_duck_objects_and_five_plush_objects_from_it_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_17/cam_high/frame_000400.jpg", "images/human_egodex/Open_the_box_remove_the_five_duck_objects_and_five_plush_objects_from_it_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_17/cam_high/frame_000400.jpg", "images/human_egodex/Open_the_box_remove_the_five_duck_objects_and_five_plush_objects_from_it_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_17/cam_high/frame_000400.jpg", "images/human_egodex/Open_the_box_remove_the_five_duck_objects_and_five_plush_objects_from_it_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_17/cam_high/frame_000427.jpg", "images/human_egodex/Open_the_box_remove_the_five_duck_objects_and_five_plush_objects_from_it_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_17/cam_high/frame_000427.jpg", "images/human_egodex/Open_the_box_remove_the_five_duck_objects_and_five_plush_objects_from_it_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_17/cam_high/frame_000427.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the box remove the five duck objects and five plush objects from it then close the box on a wooden table with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the box remove the five duck objects and five plush objects from it then close the box on a wooden table with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_utensils_from_a_wooden_table_while_sitting_against_a_white_background_episode_16-train-bf_000050-af_000400-plus-0000233", "task": "Remove utensils from a wooden table while sitting against a white background", "image": [ "images/human_egodex/Remove_utensils_from_a_wooden_table_while_sitting_against_a_white_background/episode_16/cam_high/frame_000000.jpg", "images/human_egodex/Remove_utensils_from_a_wooden_table_while_sitting_against_a_white_background/episode_16/cam_high/frame_000439.jpg", "images/human_egodex/Remove_utensils_from_a_wooden_table_while_sitting_against_a_white_background/episode_16/cam_high/frame_000050.jpg", "images/human_egodex/Remove_utensils_from_a_wooden_table_while_sitting_against_a_white_background/episode_16/cam_high/frame_000050.jpg", "images/human_egodex/Remove_utensils_from_a_wooden_table_while_sitting_against_a_white_background/episode_16/cam_high/frame_000050.jpg", "images/human_egodex/Remove_utensils_from_a_wooden_table_while_sitting_against_a_white_background/episode_16/cam_high/frame_000400.jpg", "images/human_egodex/Remove_utensils_from_a_wooden_table_while_sitting_against_a_white_background/episode_16/cam_high/frame_000400.jpg", "images/human_egodex/Remove_utensils_from_a_wooden_table_while_sitting_against_a_white_background/episode_16/cam_high/frame_000400.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove utensils from a wooden table while sitting against a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove utensils from a wooden table while sitting against a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 93.3, "hop_absolute_value": 93.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_while_sitting_at_a_table_with_a_blue_background_episode_3-train-bf_000250-af_001650-plus-0000310", "task": "Pick up individual blocks and form 3D figures on a board then shuffle the pieces while sitting at a table with a blue background", "image": [ "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_while_sitting_at_a_table_with_a_blue_background/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_while_sitting_at_a_table_with_a_blue_background/episode_3/cam_high/frame_001732.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_while_sitting_at_a_table_with_a_blue_background/episode_3/cam_high/frame_000250.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_while_sitting_at_a_table_with_a_blue_background/episode_3/cam_high/frame_000250.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_while_sitting_at_a_table_with_a_blue_background/episode_3/cam_high/frame_000250.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_while_sitting_at_a_table_with_a_blue_background/episode_3/cam_high/frame_001650.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_while_sitting_at_a_table_with_a_blue_background/episode_3/cam_high/frame_001650.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_while_sitting_at_a_table_with_a_blue_background/episode_3/cam_high/frame_001650.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up individual blocks and form 3D figures on a board then shuffle the pieces while sitting at a table with a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up individual blocks and form 3D figures on a board then shuffle the pieces while sitting at a table with a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 94.9, "hop_absolute_value": 94.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Add_a_lid_onto_a_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background_episode_2-train-bf_000175-af_000260-plus-0000102", "task": "Add a lid onto a cup placed on a wooden table while sitting against a brown background", "image": [ "images/human_egodex/Add_a_lid_onto_a_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Add_a_lid_onto_a_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_7/cam_high/frame_000337.jpg", "images/human_egodex/Add_a_lid_onto_a_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_2/cam_high/frame_000175.jpg", "images/human_egodex/Add_a_lid_onto_a_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_2/cam_high/frame_000175.jpg", "images/human_egodex/Add_a_lid_onto_a_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_2/cam_high/frame_000175.jpg", "images/human_egodex/Add_a_lid_onto_a_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_2/cam_high/frame_000260.jpg", "images/human_egodex/Add_a_lid_onto_a_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_2/cam_high/frame_000260.jpg", "images/human_egodex/Add_a_lid_onto_a_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_2/cam_high/frame_000260.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Add a lid onto a cup placed on a wooden table while sitting against a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Add a lid onto a cup placed on a wooden table while sitting against a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Use_chopsticks_to_move_cubes_and_irregular_shaped_cubes_from_one_bowl_to_another_on_a_wooden_tablecloth_while_sitting_against_a_white_background_episode_18-train-bf_000900-af_001000-plus-0000250", "task": "Use chopsticks to move cubes and irregular shaped cubes from one bowl to another on a wooden tablecloth while sitting against a white background", "image": [ "images/human_egodex/Use_chopsticks_to_move_cubes_and_irregular_shaped_cubes_from_one_bowl_to_another_on_a_wooden_tablecloth_while_sitting_against_a_white_background/episode_4/cam_high/frame_000000.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_and_irregular_shaped_cubes_from_one_bowl_to_another_on_a_wooden_tablecloth_while_sitting_against_a_white_background/episode_4/cam_high/frame_001640.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_and_irregular_shaped_cubes_from_one_bowl_to_another_on_a_wooden_tablecloth_while_sitting_against_a_white_background/episode_18/cam_high/frame_000900.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_and_irregular_shaped_cubes_from_one_bowl_to_another_on_a_wooden_tablecloth_while_sitting_against_a_white_background/episode_18/cam_high/frame_000900.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_and_irregular_shaped_cubes_from_one_bowl_to_another_on_a_wooden_tablecloth_while_sitting_against_a_white_background/episode_18/cam_high/frame_000900.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_and_irregular_shaped_cubes_from_one_bowl_to_another_on_a_wooden_tablecloth_while_sitting_against_a_white_background/episode_18/cam_high/frame_001000.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_and_irregular_shaped_cubes_from_one_bowl_to_another_on_a_wooden_tablecloth_while_sitting_against_a_white_background/episode_18/cam_high/frame_001000.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_and_irregular_shaped_cubes_from_one_bowl_to_another_on_a_wooden_tablecloth_while_sitting_against_a_white_background/episode_18/cam_high/frame_001000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use chopsticks to move cubes and irregular shaped cubes from one bowl to another on a wooden tablecloth while sitting against a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use chopsticks to move cubes and irregular shaped cubes from one bowl to another on a wooden tablecloth while sitting against a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Topple_dominoes_arranged_upright_on_a_green_table_with_a_white_background_episode_4-train-bf_000350-af_000725-plus-0000219", "task": "Topple dominoes arranged upright on a green table with a white background", "image": [ "images/human_egodex/Topple_dominoes_arranged_upright_on_a_green_table_with_a_white_background/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Topple_dominoes_arranged_upright_on_a_green_table_with_a_white_background/episode_2/cam_high/frame_001240.jpg", "images/human_egodex/Topple_dominoes_arranged_upright_on_a_green_table_with_a_white_background/episode_4/cam_high/frame_000350.jpg", "images/human_egodex/Topple_dominoes_arranged_upright_on_a_green_table_with_a_white_background/episode_4/cam_high/frame_000350.jpg", "images/human_egodex/Topple_dominoes_arranged_upright_on_a_green_table_with_a_white_background/episode_4/cam_high/frame_000350.jpg", "images/human_egodex/Topple_dominoes_arranged_upright_on_a_green_table_with_a_white_background/episode_4/cam_high/frame_000725.jpg", "images/human_egodex/Topple_dominoes_arranged_upright_on_a_green_table_with_a_white_background/episode_4/cam_high/frame_000725.jpg", "images/human_egodex/Topple_dominoes_arranged_upright_on_a_green_table_with_a_white_background/episode_4/cam_high/frame_000725.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Topple dominoes arranged upright on a green table with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Topple dominoes arranged upright on a green table with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Open_the_fridge_door_pick_up_six_food_items_from_the_table_place_them_into_the_fridge_and_close_the_fridge_door_episode_13-train-bf_000425-af_000463-plus-0000224", "task": "Open the fridge door pick up six food items from the table place them into the fridge and close the fridge door", "image": [ "images/human_egodex/Open_the_fridge_door_pick_up_six_food_items_from_the_table_place_them_into_the_fridge_and_close_the_fridge_door/episode_12/cam_high/frame_000000.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_six_food_items_from_the_table_place_them_into_the_fridge_and_close_the_fridge_door/episode_12/cam_high/frame_000657.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_six_food_items_from_the_table_place_them_into_the_fridge_and_close_the_fridge_door/episode_13/cam_high/frame_000425.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_six_food_items_from_the_table_place_them_into_the_fridge_and_close_the_fridge_door/episode_13/cam_high/frame_000425.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_six_food_items_from_the_table_place_them_into_the_fridge_and_close_the_fridge_door/episode_13/cam_high/frame_000425.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_six_food_items_from_the_table_place_them_into_the_fridge_and_close_the_fridge_door/episode_13/cam_high/frame_000463.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_six_food_items_from_the_table_place_them_into_the_fridge_and_close_the_fridge_door/episode_13/cam_high/frame_000463.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_six_food_items_from_the_table_place_them_into_the_fridge_and_close_the_fridge_door/episode_13/cam_high/frame_000463.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the fridge door pick up six food items from the table place them into the fridge and close the fridge door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the fridge door pick up six food items from the table place them into the fridge and close the fridge door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_then_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_finally_place_the_pot_and_ladle_in_the_sink_episode_6-train-bf_000000-af_000025-plus-0000319", "task": "Boil water in a pot take an egg from the carton and boil it then serve the boiled egg in a gray bowl placed on the round table and finally place the pot and ladle in the sink", "image": [ "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_then_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_finally_place_the_pot_and_ladle_in_the_sink/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_then_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_finally_place_the_pot_and_ladle_in_the_sink/episode_6/cam_high/frame_001179.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_then_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_finally_place_the_pot_and_ladle_in_the_sink/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_then_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_finally_place_the_pot_and_ladle_in_the_sink/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_then_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_finally_place_the_pot_and_ladle_in_the_sink/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_then_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_finally_place_the_pot_and_ladle_in_the_sink/episode_6/cam_high/frame_000025.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_then_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_finally_place_the_pot_and_ladle_in_the_sink/episode_6/cam_high/frame_000025.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_then_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_finally_place_the_pot_and_ladle_in_the_sink/episode_6/cam_high/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Boil water in a pot take an egg from the carton and boil it then serve the boiled egg in a gray bowl placed on the round table and finally place the pot and ladle in the sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Boil water in a pot take an egg from the carton and boil it then serve the boiled egg in a gray bowl placed on the round table and finally place the pot and ladle in the sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.1, "hop_absolute_value": 2.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Insert_twelve_blocks_into_a_dump_container_placed_on_a_blue_tablecloth_with_a_pink_background_episode_3-train-bf_000350-af_000525-plus-0000207", "task": "Insert twelve blocks into a dump container placed on a blue tablecloth with a pink background", "image": [ "images/human_egodex/Insert_twelve_blocks_into_a_dump_container_placed_on_a_blue_tablecloth_with_a_pink_background/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Insert_twelve_blocks_into_a_dump_container_placed_on_a_blue_tablecloth_with_a_pink_background/episode_0/cam_high/frame_000572.jpg", "images/human_egodex/Insert_twelve_blocks_into_a_dump_container_placed_on_a_blue_tablecloth_with_a_pink_background/episode_3/cam_high/frame_000350.jpg", "images/human_egodex/Insert_twelve_blocks_into_a_dump_container_placed_on_a_blue_tablecloth_with_a_pink_background/episode_3/cam_high/frame_000350.jpg", "images/human_egodex/Insert_twelve_blocks_into_a_dump_container_placed_on_a_blue_tablecloth_with_a_pink_background/episode_3/cam_high/frame_000350.jpg", "images/human_egodex/Insert_twelve_blocks_into_a_dump_container_placed_on_a_blue_tablecloth_with_a_pink_background/episode_3/cam_high/frame_000525.jpg", "images/human_egodex/Insert_twelve_blocks_into_a_dump_container_placed_on_a_blue_tablecloth_with_a_pink_background/episode_3/cam_high/frame_000525.jpg", "images/human_egodex/Insert_twelve_blocks_into_a_dump_container_placed_on_a_blue_tablecloth_with_a_pink_background/episode_3/cam_high/frame_000525.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert twelve blocks into a dump container placed on a blue tablecloth with a pink background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert twelve blocks into a dump container placed on a blue tablecloth with a pink background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 63.6, "hop_absolute_value": 63.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Identify_objects_such_as_red_shirts_cup_snack_wrapper_and_bottle_on_a_student_desk_with_a_pink_tablecloth_and_lavender_background_and_place_them_into_a_box_to_declutter_the_desk_episode_7-train-bf_000800-af_000900-plus-0000250", "task": "Identify objects such as red shirts cup snack wrapper and bottle on a student desk with a pink tablecloth and lavender background and place them into a box to declutter the desk", "image": [ "images/human_egodex/Identify_objects_such_as_red_shirts_cup_snack_wrapper_and_bottle_on_a_student_desk_with_a_pink_tablecloth_and_lavender_background_and_place_them_into_a_box_to_declutter_the_desk/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Identify_objects_such_as_red_shirts_cup_snack_wrapper_and_bottle_on_a_student_desk_with_a_pink_tablecloth_and_lavender_background_and_place_them_into_a_box_to_declutter_the_desk/episode_7/cam_high/frame_000930.jpg", "images/human_egodex/Identify_objects_such_as_red_shirts_cup_snack_wrapper_and_bottle_on_a_student_desk_with_a_pink_tablecloth_and_lavender_background_and_place_them_into_a_box_to_declutter_the_desk/episode_7/cam_high/frame_000800.jpg", "images/human_egodex/Identify_objects_such_as_red_shirts_cup_snack_wrapper_and_bottle_on_a_student_desk_with_a_pink_tablecloth_and_lavender_background_and_place_them_into_a_box_to_declutter_the_desk/episode_7/cam_high/frame_000800.jpg", "images/human_egodex/Identify_objects_such_as_red_shirts_cup_snack_wrapper_and_bottle_on_a_student_desk_with_a_pink_tablecloth_and_lavender_background_and_place_them_into_a_box_to_declutter_the_desk/episode_7/cam_high/frame_000800.jpg", "images/human_egodex/Identify_objects_such_as_red_shirts_cup_snack_wrapper_and_bottle_on_a_student_desk_with_a_pink_tablecloth_and_lavender_background_and_place_them_into_a_box_to_declutter_the_desk/episode_7/cam_high/frame_000900.jpg", "images/human_egodex/Identify_objects_such_as_red_shirts_cup_snack_wrapper_and_bottle_on_a_student_desk_with_a_pink_tablecloth_and_lavender_background_and_place_them_into_a_box_to_declutter_the_desk/episode_7/cam_high/frame_000900.jpg", "images/human_egodex/Identify_objects_such_as_red_shirts_cup_snack_wrapper_and_bottle_on_a_student_desk_with_a_pink_tablecloth_and_lavender_background_and_place_them_into_a_box_to_declutter_the_desk/episode_7/cam_high/frame_000900.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Identify objects such as red shirts cup snack wrapper and bottle on a student desk with a pink tablecloth and lavender background and place them into a box to declutter the desk`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Identify objects such as red shirts cup snack wrapper and bottle on a student desk with a pink tablecloth and lavender background and place them into a box to declutter the desk`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_rubber_band_from_around_the_group_of_small_objects_like_cards_episode_18-train-bf_000000-af_000125-plus-0000092", "task": "Remove the rubber band from around the group of small objects like cards", "image": [ "images/human_egodex/Remove_the_rubber_band_from_around_the_group_of_small_objects_like_cards/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_rubber_band_from_around_the_group_of_small_objects_like_cards/episode_18/cam_high/frame_000294.jpg", "images/human_egodex/Remove_the_rubber_band_from_around_the_group_of_small_objects_like_cards/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_rubber_band_from_around_the_group_of_small_objects_like_cards/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_rubber_band_from_around_the_group_of_small_objects_like_cards/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_rubber_band_from_around_the_group_of_small_objects_like_cards/episode_18/cam_high/frame_000125.jpg", "images/human_egodex/Remove_the_rubber_band_from_around_the_group_of_small_objects_like_cards/episode_18/cam_high/frame_000125.jpg", "images/human_egodex/Remove_the_rubber_band_from_around_the_group_of_small_objects_like_cards/episode_18/cam_high/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the rubber band from around the group of small objects like cards`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the rubber band from around the group of small objects like cards`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 45.5, "hop_absolute_value": 45.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Braid_the_three_strands_of_ribbon_into_a_braid_while_sitting_against_a_brown_background_episode_0-train-bf_000075-af_000150-plus-0000388", "task": "Braid the three strands of ribbon into a braid while sitting against a brown background", "image": [ "images/human_egodex/Braid_the_three_strands_of_ribbon_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000683.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000075.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000075.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000075.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000150.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000150.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Braid the three strands of ribbon into a braid while sitting against a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Braid the three strands of ribbon into a braid while sitting against a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Arrange_dominoes_upright_in_a_pattern_on_a_green_table_episode_13-train-bf_000550-af_000575-plus-0000358", "task": "Arrange dominoes upright in a pattern on a green table", "image": [ "images/human_egodex/Arrange_dominoes_upright_in_a_pattern_on_a_green_table/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Arrange_dominoes_upright_in_a_pattern_on_a_green_table/episode_0/cam_high/frame_004783.jpg", "images/human_egodex/Arrange_dominoes_upright_in_a_pattern_on_a_green_table/episode_13/cam_high/frame_000550.jpg", "images/human_egodex/Arrange_dominoes_upright_in_a_pattern_on_a_green_table/episode_13/cam_high/frame_000550.jpg", "images/human_egodex/Arrange_dominoes_upright_in_a_pattern_on_a_green_table/episode_13/cam_high/frame_000550.jpg", "images/human_egodex/Arrange_dominoes_upright_in_a_pattern_on_a_green_table/episode_13/cam_high/frame_000575.jpg", "images/human_egodex/Arrange_dominoes_upright_in_a_pattern_on_a_green_table/episode_13/cam_high/frame_000575.jpg", "images/human_egodex/Arrange_dominoes_upright_in_a_pattern_on_a_green_table/episode_13/cam_high/frame_000575.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Arrange dominoes upright in a pattern on a green table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Arrange dominoes upright in a pattern on a green table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.5, "hop_absolute_value": 1.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_from_the_red_tablecloth_and_fold_it_back_up_episode_15-train-bf_000075-af_000725-plus-0000232", "task": "Remove the blue and yellow plates forks spoons knives and cups from the red tablecloth and fold it back up", "image": [ "images/human_egodex/Remove_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_from_the_red_tablecloth_and_fold_it_back_up/episode_15/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_from_the_red_tablecloth_and_fold_it_back_up/episode_15/cam_high/frame_001144.jpg", "images/human_egodex/Remove_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_from_the_red_tablecloth_and_fold_it_back_up/episode_15/cam_high/frame_000075.jpg", "images/human_egodex/Remove_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_from_the_red_tablecloth_and_fold_it_back_up/episode_15/cam_high/frame_000075.jpg", "images/human_egodex/Remove_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_from_the_red_tablecloth_and_fold_it_back_up/episode_15/cam_high/frame_000075.jpg", "images/human_egodex/Remove_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_from_the_red_tablecloth_and_fold_it_back_up/episode_15/cam_high/frame_000725.jpg", "images/human_egodex/Remove_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_from_the_red_tablecloth_and_fold_it_back_up/episode_15/cam_high/frame_000725.jpg", "images/human_egodex/Remove_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_from_the_red_tablecloth_and_fold_it_back_up/episode_15/cam_high/frame_000725.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the blue and yellow plates forks spoons knives and cups from the red tablecloth and fold it back up`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the blue and yellow plates forks spoons knives and cups from the red tablecloth and fold it back up`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 61.9, "hop_absolute_value": 61.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_egg_from_bowl_return_it_to_egg_carton_empty_pot_and_ladle_from_sink_and_place_all_objects_back_to_their_original_positions_episode_2-train-bf_000075-af_000675-plus-0000253", "task": "Remove egg from bowl return it to egg carton empty pot and ladle from sink and place all objects back to their original positions", "image": [ "images/human_egodex/Remove_egg_from_bowl_return_it_to_egg_carton_empty_pot_and_ladle_from_sink_and_place_all_objects_back_to_their_original_positions/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Remove_egg_from_bowl_return_it_to_egg_carton_empty_pot_and_ladle_from_sink_and_place_all_objects_back_to_their_original_positions/episode_2/cam_high/frame_000871.jpg", "images/human_egodex/Remove_egg_from_bowl_return_it_to_egg_carton_empty_pot_and_ladle_from_sink_and_place_all_objects_back_to_their_original_positions/episode_2/cam_high/frame_000075.jpg", "images/human_egodex/Remove_egg_from_bowl_return_it_to_egg_carton_empty_pot_and_ladle_from_sink_and_place_all_objects_back_to_their_original_positions/episode_2/cam_high/frame_000075.jpg", "images/human_egodex/Remove_egg_from_bowl_return_it_to_egg_carton_empty_pot_and_ladle_from_sink_and_place_all_objects_back_to_their_original_positions/episode_2/cam_high/frame_000075.jpg", "images/human_egodex/Remove_egg_from_bowl_return_it_to_egg_carton_empty_pot_and_ladle_from_sink_and_place_all_objects_back_to_their_original_positions/episode_2/cam_high/frame_000675.jpg", "images/human_egodex/Remove_egg_from_bowl_return_it_to_egg_carton_empty_pot_and_ladle_from_sink_and_place_all_objects_back_to_their_original_positions/episode_2/cam_high/frame_000675.jpg", "images/human_egodex/Remove_egg_from_bowl_return_it_to_egg_carton_empty_pot_and_ladle_from_sink_and_place_all_objects_back_to_their_original_positions/episode_2/cam_high/frame_000675.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove egg from bowl return it to egg carton empty pot and ladle from sink and place all objects back to their original positions`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove egg from bowl return it to egg carton empty pot and ladle from sink and place all objects back to their original positions`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 77.4, "hop_absolute_value": 77.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_clear_tupperware_from_the_shelf_with_the_right_hand_and_place_it_back_on_the_wooden_table_episode_1-train-bf_000325-af_000425-plus-0000164", "task": "Remove the clear tupperware from the shelf with the right hand and place it back on the wooden table", "image": [ "images/human_egodex/Remove_the_clear_tupperware_from_the_shelf_with_the_right_hand_and_place_it_back_on_the_wooden_table/episode_1/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_clear_tupperware_from_the_shelf_with_the_right_hand_and_place_it_back_on_the_wooden_table/episode_1/cam_high/frame_000595.jpg", "images/human_egodex/Remove_the_clear_tupperware_from_the_shelf_with_the_right_hand_and_place_it_back_on_the_wooden_table/episode_1/cam_high/frame_000325.jpg", "images/human_egodex/Remove_the_clear_tupperware_from_the_shelf_with_the_right_hand_and_place_it_back_on_the_wooden_table/episode_1/cam_high/frame_000325.jpg", "images/human_egodex/Remove_the_clear_tupperware_from_the_shelf_with_the_right_hand_and_place_it_back_on_the_wooden_table/episode_1/cam_high/frame_000325.jpg", "images/human_egodex/Remove_the_clear_tupperware_from_the_shelf_with_the_right_hand_and_place_it_back_on_the_wooden_table/episode_1/cam_high/frame_000425.jpg", "images/human_egodex/Remove_the_clear_tupperware_from_the_shelf_with_the_right_hand_and_place_it_back_on_the_wooden_table/episode_1/cam_high/frame_000425.jpg", "images/human_egodex/Remove_the_clear_tupperware_from_the_shelf_with_the_right_hand_and_place_it_back_on_the_wooden_table/episode_1/cam_high/frame_000425.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the clear tupperware from the shelf with the right hand and place it back on the wooden table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the clear tupperware from the shelf with the right hand and place it back on the wooden table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Flatten_the_crumpled_paper_on_a_pink_tablecloth_with_a_lavender_background_episode_8-train-bf_000150-af_000400-plus-0000164", "task": "Flatten the crumpled paper on a pink tablecloth with a lavender background", "image": [ "images/human_egodex/Flatten_the_crumpled_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Flatten_the_crumpled_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_8/cam_high/frame_001102.jpg", "images/human_egodex/Flatten_the_crumpled_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_8/cam_high/frame_000150.jpg", "images/human_egodex/Flatten_the_crumpled_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_8/cam_high/frame_000150.jpg", "images/human_egodex/Flatten_the_crumpled_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_8/cam_high/frame_000150.jpg", "images/human_egodex/Flatten_the_crumpled_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_8/cam_high/frame_000400.jpg", "images/human_egodex/Flatten_the_crumpled_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_8/cam_high/frame_000400.jpg", "images/human_egodex/Flatten_the_crumpled_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_8/cam_high/frame_000400.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Flatten the crumpled paper on a pink tablecloth with a lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Flatten the crumpled paper on a pink tablecloth with a lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 26.3, "hop_absolute_value": 26.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting_episode_7-train-bf_000275-af_000350-plus-0000446", "task": "Disassemble tiles on a metal table with a lavender background while sitting", "image": [ "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_0/cam_high/frame_000966.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_7/cam_high/frame_000275.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_7/cam_high/frame_000275.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_7/cam_high/frame_000275.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_7/cam_high/frame_000350.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_7/cam_high/frame_000350.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_7/cam_high/frame_000350.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble tiles on a metal table with a lavender background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble tiles on a metal table with a lavender background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Stack_the_removed_block_back_onto_the_stack_of_Jenga_blocks_on_the_wooden_table_episode_18-train-bf_000000-af_000125-plus-0000035", "task": "Stack the removed block back onto the stack of Jenga blocks on the wooden table", "image": [ "images/human_egodex/Stack_the_removed_block_back_onto_the_stack_of_Jenga_blocks_on_the_wooden_table/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Stack_the_removed_block_back_onto_the_stack_of_Jenga_blocks_on_the_wooden_table/episode_18/cam_high/frame_000166.jpg", "images/human_egodex/Stack_the_removed_block_back_onto_the_stack_of_Jenga_blocks_on_the_wooden_table/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Stack_the_removed_block_back_onto_the_stack_of_Jenga_blocks_on_the_wooden_table/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Stack_the_removed_block_back_onto_the_stack_of_Jenga_blocks_on_the_wooden_table/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Stack_the_removed_block_back_onto_the_stack_of_Jenga_blocks_on_the_wooden_table/episode_18/cam_high/frame_000125.jpg", "images/human_egodex/Stack_the_removed_block_back_onto_the_stack_of_Jenga_blocks_on_the_wooden_table/episode_18/cam_high/frame_000125.jpg", "images/human_egodex/Stack_the_removed_block_back_onto_the_stack_of_Jenga_blocks_on_the_wooden_table/episode_18/cam_high/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Stack the removed block back onto the stack of Jenga blocks on the wooden table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Stack the removed block back onto the stack of Jenga blocks on the wooden table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_around_episode_8-train-bf_000875-af_001125-plus-0000172", "task": "Pick up individual blocks and form 3D figures on a board then shuffle the pieces around", "image": [ "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_around/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_around/episode_3/cam_high/frame_001896.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_around/episode_8/cam_high/frame_000875.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_around/episode_8/cam_high/frame_000875.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_around/episode_8/cam_high/frame_000875.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_around/episode_8/cam_high/frame_001125.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_around/episode_8/cam_high/frame_001125.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_then_shuffle_the_pieces_around/episode_8/cam_high/frame_001125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up individual blocks and form 3D figures on a board then shuffle the pieces around`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up individual blocks and form 3D figures on a board then shuffle the pieces around`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 21.7, "hop_absolute_value": 21.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_the_assembled_puzzle4_and_place_the_pieces_back_on_the_wooden_table_episode_13-train-bf_000275-af_000350-plus-0000307", "task": "Disassemble the assembled puzzle4 and place the pieces back on the wooden table", "image": [ "images/human_egodex/Disassemble_the_assembled_puzzle4_and_place_the_pieces_back_on_the_wooden_table/episode_13/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle4_and_place_the_pieces_back_on_the_wooden_table/episode_13/cam_high/frame_000469.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle4_and_place_the_pieces_back_on_the_wooden_table/episode_13/cam_high/frame_000275.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle4_and_place_the_pieces_back_on_the_wooden_table/episode_13/cam_high/frame_000275.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle4_and_place_the_pieces_back_on_the_wooden_table/episode_13/cam_high/frame_000275.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle4_and_place_the_pieces_back_on_the_wooden_table/episode_13/cam_high/frame_000350.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle4_and_place_the_pieces_back_on_the_wooden_table/episode_13/cam_high/frame_000350.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle4_and_place_the_pieces_back_on_the_wooden_table/episode_13/cam_high/frame_000350.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble the assembled puzzle4 and place the pieces back on the wooden table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble the assembled puzzle4 and place the pieces back on the wooden table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 42.9, "hop_absolute_value": 42.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_a_croissant_from_the_blue_bowl_and_place_it_in_the_green_bowl_episode_0-train-bf_000025-af_000100-plus-0000023", "task": "Pick up a croissant from the blue bowl and place it in the green bowl", "image": [ "images/human_egodex/Pick_up_a_croissant_from_the_blue_bowl_and_place_it_in_the_green_bowl/episode_19/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_a_croissant_from_the_blue_bowl_and_place_it_in_the_green_bowl/episode_19/cam_high/frame_000151.jpg", "images/human_egodex/Pick_up_a_croissant_from_the_blue_bowl_and_place_it_in_the_green_bowl/episode_0/cam_high/frame_000025.jpg", "images/human_egodex/Pick_up_a_croissant_from_the_blue_bowl_and_place_it_in_the_green_bowl/episode_0/cam_high/frame_000025.jpg", "images/human_egodex/Pick_up_a_croissant_from_the_blue_bowl_and_place_it_in_the_green_bowl/episode_0/cam_high/frame_000025.jpg", "images/human_egodex/Pick_up_a_croissant_from_the_blue_bowl_and_place_it_in_the_green_bowl/episode_0/cam_high/frame_000100.jpg", "images/human_egodex/Pick_up_a_croissant_from_the_blue_bowl_and_place_it_in_the_green_bowl/episode_0/cam_high/frame_000100.jpg", "images/human_egodex/Pick_up_a_croissant_from_the_blue_bowl_and_place_it_in_the_green_bowl/episode_0/cam_high/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up a croissant from the blue bowl and place it in the green bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up a croissant from the blue bowl and place it in the green bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unstake_the_lego_bricks_to_disassemble_the_structure_episode_3-train-bf_000175-af_000235-plus-0000082", "task": "Unstake the lego bricks to disassemble the structure", "image": [ "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_structure/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_structure/episode_18/cam_high/frame_000276.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_structure/episode_3/cam_high/frame_000175.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_structure/episode_3/cam_high/frame_000175.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_structure/episode_3/cam_high/frame_000175.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_structure/episode_3/cam_high/frame_000235.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_structure/episode_3/cam_high/frame_000235.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_structure/episode_3/cam_high/frame_000235.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unstake the lego bricks to disassemble the structure`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unstake the lego bricks to disassemble the structure`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Insert_a_tennis_ball_among_three_tennis_balls_episode_4-train-bf_000250-af_000300-plus-0000370", "task": "Insert a tennis ball among three tennis balls", "image": [ "images/human_egodex/Insert_a_tennis_ball_among_three_tennis_balls/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Insert_a_tennis_ball_among_three_tennis_balls/episode_9/cam_high/frame_000850.jpg", "images/human_egodex/Insert_a_tennis_ball_among_three_tennis_balls/episode_4/cam_high/frame_000250.jpg", "images/human_egodex/Insert_a_tennis_ball_among_three_tennis_balls/episode_4/cam_high/frame_000250.jpg", "images/human_egodex/Insert_a_tennis_ball_among_three_tennis_balls/episode_4/cam_high/frame_000250.jpg", "images/human_egodex/Insert_a_tennis_ball_among_three_tennis_balls/episode_4/cam_high/frame_000300.jpg", "images/human_egodex/Insert_a_tennis_ball_among_three_tennis_balls/episode_4/cam_high/frame_000300.jpg", "images/human_egodex/Insert_a_tennis_ball_among_three_tennis_balls/episode_4/cam_high/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert a tennis ball among three tennis balls`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert a tennis ball among three tennis balls`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.5, "hop_absolute_value": 9.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unwrap_a_chickenburger_from_plastic_wrap_on_a_pink_tablecloth_with_a_lavender_background_episode_18-train-bf_000025-af_000300-plus-0000175", "task": "Unwrap a chickenburger from plastic wrap on a pink tablecloth with a lavender background", "image": [ "images/human_egodex/Unwrap_a_chickenburger_from_plastic_wrap_on_a_pink_tablecloth_with_a_lavender_background/episode_12/cam_high/frame_000000.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_plastic_wrap_on_a_pink_tablecloth_with_a_lavender_background/episode_12/cam_high/frame_000738.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_plastic_wrap_on_a_pink_tablecloth_with_a_lavender_background/episode_18/cam_high/frame_000025.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_plastic_wrap_on_a_pink_tablecloth_with_a_lavender_background/episode_18/cam_high/frame_000025.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_plastic_wrap_on_a_pink_tablecloth_with_a_lavender_background/episode_18/cam_high/frame_000025.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_plastic_wrap_on_a_pink_tablecloth_with_a_lavender_background/episode_18/cam_high/frame_000300.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_plastic_wrap_on_a_pink_tablecloth_with_a_lavender_background/episode_18/cam_high/frame_000300.jpg", "images/human_egodex/Unwrap_a_chickenburger_from_plastic_wrap_on_a_pink_tablecloth_with_a_lavender_background/episode_18/cam_high/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unwrap a chickenburger from plastic wrap on a pink tablecloth with a lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unwrap a chickenburger from plastic wrap on a pink tablecloth with a lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 78.6, "hop_absolute_value": 78.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Braid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background_episode_19-train-bf_000125-af_000225-plus-0000390", "task": "Braid the three strands of ribbon while sitting against a brown background", "image": [ "images/human_egodex/Braid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_18/cam_high/frame_000717.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_19/cam_high/frame_000125.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_19/cam_high/frame_000125.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_19/cam_high/frame_000125.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_19/cam_high/frame_000225.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_19/cam_high/frame_000225.jpg", "images/human_egodex/Braid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_19/cam_high/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Braid the three strands of ribbon while sitting against a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Braid the three strands of ribbon while sitting against a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_the_leg_from_the_stool_on_a_wooden_table_with_a_red_background_while_sitting_episode_17-train-bf_000350-af_000450-plus-0000177", "task": "Disassemble the leg from the stool on a wooden table with a red background while sitting", "image": [ "images/human_egodex/Disassemble_the_leg_from_the_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_the_leg_from_the_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_7/cam_high/frame_000250.jpg", "images/human_egodex/Disassemble_the_leg_from_the_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_17/cam_high/frame_000350.jpg", "images/human_egodex/Disassemble_the_leg_from_the_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_17/cam_high/frame_000350.jpg", "images/human_egodex/Disassemble_the_leg_from_the_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_17/cam_high/frame_000350.jpg", "images/human_egodex/Disassemble_the_leg_from_the_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_17/cam_high/frame_000450.jpg", "images/human_egodex/Disassemble_the_leg_from_the_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_17/cam_high/frame_000450.jpg", "images/human_egodex/Disassemble_the_leg_from_the_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_17/cam_high/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble the leg from the stool on a wooden table with a red background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble the leg from the stool on a wooden table with a red background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.1, "hop_absolute_value": 57.1, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink_episode_8-train-bf_000225-af_000875-plus-0000299", "task": "Boil water in a pot take an egg from the carton and boil it serve the boiled egg in a gray bowl placed on the round table and place the pot and ladle in the sink", "image": [ "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_000931.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_000225.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_000225.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_000225.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_000875.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_000875.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_000875.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Boil water in a pot take an egg from the carton and boil it serve the boiled egg in a gray bowl placed on the round table and place the pot and ladle in the sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Boil water in a pot take an egg from the carton and boil it serve the boiled egg in a gray bowl placed on the round table and place the pot and ladle in the sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 92.9, "hop_absolute_value": 92.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Assemble_four_structures_on_a_wooden_table_while_sitting_against_a_white_background_episode_9-train-bf_000025-af_000150-plus-0000124", "task": "Assemble four structures on a wooden table while sitting against a white background", "image": [ "images/human_egodex/Assemble_four_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Assemble_four_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_18/cam_high/frame_000501.jpg", "images/human_egodex/Assemble_four_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_9/cam_high/frame_000025.jpg", "images/human_egodex/Assemble_four_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_9/cam_high/frame_000025.jpg", "images/human_egodex/Assemble_four_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_9/cam_high/frame_000025.jpg", "images/human_egodex/Assemble_four_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_9/cam_high/frame_000150.jpg", "images/human_egodex/Assemble_four_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_9/cam_high/frame_000150.jpg", "images/human_egodex/Assemble_four_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_9/cam_high/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Assemble four structures on a wooden table while sitting against a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Assemble four structures on a wooden table while sitting against a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 35.7, "hop_absolute_value": 35.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unwrap_the_croissant_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate_episode_8-train-bf_000450-af_000600-plus-0000213", "task": "Unwrap the croissant from the plastic wrap and place it back on the gray plate", "image": [ "images/human_egodex/Unwrap_the_croissant_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Unwrap_the_croissant_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000688.jpg", "images/human_egodex/Unwrap_the_croissant_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000450.jpg", "images/human_egodex/Unwrap_the_croissant_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000450.jpg", "images/human_egodex/Unwrap_the_croissant_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000450.jpg", "images/human_egodex/Unwrap_the_croissant_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000600.jpg", "images/human_egodex/Unwrap_the_croissant_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000600.jpg", "images/human_egodex/Unwrap_the_croissant_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000600.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unwrap the croissant from the plastic wrap and place it back on the gray plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unwrap the croissant from the plastic wrap and place it back on the gray plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Add_lids_onto_four_cups_placed_on_a_wooden_table_with_a_red_background_episode_16-train-bf_000050-af_000250-plus-0000162", "task": "Add lids onto four cups placed on a wooden table with a red background", "image": [ "images/human_egodex/Add_lids_onto_four_cups_placed_on_a_wooden_table_with_a_red_background/episode_16/cam_high/frame_000000.jpg", "images/human_egodex/Add_lids_onto_four_cups_placed_on_a_wooden_table_with_a_red_background/episode_16/cam_high/frame_000390.jpg", "images/human_egodex/Add_lids_onto_four_cups_placed_on_a_wooden_table_with_a_red_background/episode_16/cam_high/frame_000050.jpg", "images/human_egodex/Add_lids_onto_four_cups_placed_on_a_wooden_table_with_a_red_background/episode_16/cam_high/frame_000050.jpg", "images/human_egodex/Add_lids_onto_four_cups_placed_on_a_wooden_table_with_a_red_background/episode_16/cam_high/frame_000050.jpg", "images/human_egodex/Add_lids_onto_four_cups_placed_on_a_wooden_table_with_a_red_background/episode_16/cam_high/frame_000250.jpg", "images/human_egodex/Add_lids_onto_four_cups_placed_on_a_wooden_table_with_a_red_background/episode_16/cam_high/frame_000250.jpg", "images/human_egodex/Add_lids_onto_four_cups_placed_on_a_wooden_table_with_a_red_background/episode_16/cam_high/frame_000250.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Add lids onto four cups placed on a wooden table with a red background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Add lids onto four cups placed on a wooden table with a red background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 61.5, "hop_absolute_value": 61.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Clean_up_chess_pieces_from_the_chessboard_placed_on_a_wooden_table_episode_4-train-bf_000325-af_001000-plus-0000222", "task": "Clean up chess pieces from the chessboard placed on a wooden table", "image": [ "images/human_egodex/Clean_up_chess_pieces_from_the_chessboard_placed_on_a_wooden_table/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Clean_up_chess_pieces_from_the_chessboard_placed_on_a_wooden_table/episode_8/cam_high/frame_001541.jpg", "images/human_egodex/Clean_up_chess_pieces_from_the_chessboard_placed_on_a_wooden_table/episode_4/cam_high/frame_000325.jpg", "images/human_egodex/Clean_up_chess_pieces_from_the_chessboard_placed_on_a_wooden_table/episode_4/cam_high/frame_000325.jpg", "images/human_egodex/Clean_up_chess_pieces_from_the_chessboard_placed_on_a_wooden_table/episode_4/cam_high/frame_000325.jpg", "images/human_egodex/Clean_up_chess_pieces_from_the_chessboard_placed_on_a_wooden_table/episode_4/cam_high/frame_001000.jpg", "images/human_egodex/Clean_up_chess_pieces_from_the_chessboard_placed_on_a_wooden_table/episode_4/cam_high/frame_001000.jpg", "images/human_egodex/Clean_up_chess_pieces_from_the_chessboard_placed_on_a_wooden_table/episode_4/cam_high/frame_001000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Clean up chess pieces from the chessboard placed on a wooden table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Clean up chess pieces from the chessboard placed on a wooden table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_four_legos_while_sitting_at_a_green_tablecloth_with_a_blue_background_episode_6-train-bf_000000-af_000234-plus-0000088", "task": "Disassemble four legos while sitting at a green tablecloth with a blue background", "image": [ "images/human_egodex/Disassemble_four_legos_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_four_legos_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000234.jpg", "images/human_egodex/Disassemble_four_legos_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_four_legos_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_four_legos_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_four_legos_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000234.jpg", "images/human_egodex/Disassemble_four_legos_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000234.jpg", "images/human_egodex/Disassemble_four_legos_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000234.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble four legos while sitting at a green tablecloth with a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble four legos while sitting at a green tablecloth with a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_blue_plate_blue_cup_blue_fork_blue_spoon_and_blue_knife_from_the_red_tablecloth_then_fold_the_tablecloth_episode_10-train-bf_000675-af_000900-plus-0000204", "task": "Remove the blue plate blue cup blue fork blue spoon and blue knife from the red tablecloth then fold the tablecloth", "image": [ "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_blue_spoon_and_blue_knife_from_the_red_tablecloth_then_fold_the_tablecloth/episode_10/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_blue_spoon_and_blue_knife_from_the_red_tablecloth_then_fold_the_tablecloth/episode_10/cam_high/frame_001114.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_blue_spoon_and_blue_knife_from_the_red_tablecloth_then_fold_the_tablecloth/episode_10/cam_high/frame_000675.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_blue_spoon_and_blue_knife_from_the_red_tablecloth_then_fold_the_tablecloth/episode_10/cam_high/frame_000675.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_blue_spoon_and_blue_knife_from_the_red_tablecloth_then_fold_the_tablecloth/episode_10/cam_high/frame_000675.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_blue_spoon_and_blue_knife_from_the_red_tablecloth_then_fold_the_tablecloth/episode_10/cam_high/frame_000900.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_blue_spoon_and_blue_knife_from_the_red_tablecloth_then_fold_the_tablecloth/episode_10/cam_high/frame_000900.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_blue_spoon_and_blue_knife_from_the_red_tablecloth_then_fold_the_tablecloth/episode_10/cam_high/frame_000900.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the blue plate blue cup blue fork blue spoon and blue knife from the red tablecloth then fold the tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the blue plate blue cup blue fork blue spoon and blue knife from the red tablecloth then fold the tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 52.9, "hop_absolute_value": 52.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background_episode_12-train-bf_000050-af_000675-plus-0000283", "task": "Remove the USB from the hairbrush using the left hand while sitting against a blue background", "image": [ "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_11/cam_high/frame_001049.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000050.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000050.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000050.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000675.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000675.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000675.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the USB from the hairbrush using the left hand while sitting against a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the USB from the hairbrush using the left hand while sitting against a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 96.2, "hop_absolute_value": 96.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Assemble_ten_tiles_on_a_wooden_table_while_sitting_against_a_white_background_episode_17-train-bf_000250-af_000300-plus-0000362", "task": "Assemble ten tiles on a wooden table while sitting against a white background", "image": [ "images/human_egodex/Assemble_ten_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Assemble_ten_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_2/cam_high/frame_001270.jpg", "images/human_egodex/Assemble_ten_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_17/cam_high/frame_000250.jpg", "images/human_egodex/Assemble_ten_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_17/cam_high/frame_000250.jpg", "images/human_egodex/Assemble_ten_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_17/cam_high/frame_000250.jpg", "images/human_egodex/Assemble_ten_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_17/cam_high/frame_000300.jpg", "images/human_egodex/Assemble_ten_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_17/cam_high/frame_000300.jpg", "images/human_egodex/Assemble_ten_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_17/cam_high/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Assemble ten tiles on a wooden table while sitting against a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Assemble ten tiles on a wooden table while sitting against a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 8.0, "hop_absolute_value": 8.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_USB_device_from_the_USB_port_while_sitting_at_a_table_with_a_white_background_episode_15-train-bf_000175-af_000400-plus-0000200", "task": "Remove the USB device from the USB port while sitting at a table with a white background", "image": [ "images/human_egodex/Remove_the_USB_device_from_the_USB_port_while_sitting_at_a_table_with_a_white_background/episode_17/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_USB_device_from_the_USB_port_while_sitting_at_a_table_with_a_white_background/episode_17/cam_high/frame_000484.jpg", "images/human_egodex/Remove_the_USB_device_from_the_USB_port_while_sitting_at_a_table_with_a_white_background/episode_15/cam_high/frame_000175.jpg", "images/human_egodex/Remove_the_USB_device_from_the_USB_port_while_sitting_at_a_table_with_a_white_background/episode_15/cam_high/frame_000175.jpg", "images/human_egodex/Remove_the_USB_device_from_the_USB_port_while_sitting_at_a_table_with_a_white_background/episode_15/cam_high/frame_000175.jpg", "images/human_egodex/Remove_the_USB_device_from_the_USB_port_while_sitting_at_a_table_with_a_white_background/episode_15/cam_high/frame_000400.jpg", "images/human_egodex/Remove_the_USB_device_from_the_USB_port_while_sitting_at_a_table_with_a_white_background/episode_15/cam_high/frame_000400.jpg", "images/human_egodex/Remove_the_USB_device_from_the_USB_port_while_sitting_at_a_table_with_a_white_background/episode_15/cam_high/frame_000400.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the USB device from the USB port while sitting at a table with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the USB device from the USB port while sitting at a table with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 69.2, "hop_absolute_value": 69.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_blue_plate_blue_cup_blue_fork_and_banana_from_the_red_tablecloth_and_fold_it_back_up_episode_11-train-bf_000025-af_000225-plus-0000171", "task": "Remove the blue plate blue cup blue fork and banana from the red tablecloth and fold it back up", "image": [ "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_and_banana_from_the_red_tablecloth_and_fold_it_back_up/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_and_banana_from_the_red_tablecloth_and_fold_it_back_up/episode_11/cam_high/frame_000587.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_and_banana_from_the_red_tablecloth_and_fold_it_back_up/episode_11/cam_high/frame_000025.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_and_banana_from_the_red_tablecloth_and_fold_it_back_up/episode_11/cam_high/frame_000025.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_and_banana_from_the_red_tablecloth_and_fold_it_back_up/episode_11/cam_high/frame_000025.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_and_banana_from_the_red_tablecloth_and_fold_it_back_up/episode_11/cam_high/frame_000225.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_and_banana_from_the_red_tablecloth_and_fold_it_back_up/episode_11/cam_high/frame_000225.jpg", "images/human_egodex/Remove_the_blue_plate_blue_cup_blue_fork_and_banana_from_the_red_tablecloth_and_fold_it_back_up/episode_11/cam_high/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the blue plate blue cup blue fork and banana from the red tablecloth and fold it back up`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the blue plate blue cup blue fork and banana from the red tablecloth and fold it back up`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 36.4, "hop_absolute_value": 36.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_the_assembled_puzzle_and_place_the_pieces_back_on_the_wooden_table_episode_8-train-bf_000125-af_000275-plus-0000154", "task": "Disassemble the assembled puzzle and place the pieces back on the wooden table", "image": [ "images/human_egodex/Disassemble_the_assembled_puzzle_and_place_the_pieces_back_on_the_wooden_table/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle_and_place_the_pieces_back_on_the_wooden_table/episode_8/cam_high/frame_000606.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle_and_place_the_pieces_back_on_the_wooden_table/episode_8/cam_high/frame_000125.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle_and_place_the_pieces_back_on_the_wooden_table/episode_8/cam_high/frame_000125.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle_and_place_the_pieces_back_on_the_wooden_table/episode_8/cam_high/frame_000125.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle_and_place_the_pieces_back_on_the_wooden_table/episode_8/cam_high/frame_000275.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle_and_place_the_pieces_back_on_the_wooden_table/episode_8/cam_high/frame_000275.jpg", "images/human_egodex/Disassemble_the_assembled_puzzle_and_place_the_pieces_back_on_the_wooden_table/episode_8/cam_high/frame_000275.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble the assembled puzzle and place the pieces back on the wooden table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble the assembled puzzle and place the pieces back on the wooden table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 31.6, "hop_absolute_value": 31.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_yellow_rubber_duck_from_the_silver_bowl_with_the_left_hand_and_place_it_into_the_turquoise_basket_episode_4-train-bf_000100-af_000150-plus-0000088", "task": "Pick up the yellow rubber duck from the silver bowl with the left hand and place it into the turquoise basket", "image": [ "images/human_egodex/Pick_up_the_yellow_rubber_duck_from_the_silver_bowl_with_the_left_hand_and_place_it_into_the_turquoise_basket/episode_4/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_yellow_rubber_duck_from_the_silver_bowl_with_the_left_hand_and_place_it_into_the_turquoise_basket/episode_4/cam_high/frame_000300.jpg", "images/human_egodex/Pick_up_the_yellow_rubber_duck_from_the_silver_bowl_with_the_left_hand_and_place_it_into_the_turquoise_basket/episode_4/cam_high/frame_000100.jpg", "images/human_egodex/Pick_up_the_yellow_rubber_duck_from_the_silver_bowl_with_the_left_hand_and_place_it_into_the_turquoise_basket/episode_4/cam_high/frame_000100.jpg", "images/human_egodex/Pick_up_the_yellow_rubber_duck_from_the_silver_bowl_with_the_left_hand_and_place_it_into_the_turquoise_basket/episode_4/cam_high/frame_000100.jpg", "images/human_egodex/Pick_up_the_yellow_rubber_duck_from_the_silver_bowl_with_the_left_hand_and_place_it_into_the_turquoise_basket/episode_4/cam_high/frame_000150.jpg", "images/human_egodex/Pick_up_the_yellow_rubber_duck_from_the_silver_bowl_with_the_left_hand_and_place_it_into_the_turquoise_basket/episode_4/cam_high/frame_000150.jpg", "images/human_egodex/Pick_up_the_yellow_rubber_duck_from_the_silver_bowl_with_the_left_hand_and_place_it_into_the_turquoise_basket/episode_4/cam_high/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the yellow rubber duck from the silver bowl with the left hand and place it into the turquoise basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the yellow rubber duck from the silver bowl with the left hand and place it into the turquoise basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Grab_the_croissant_from_the_gray_plate_and_wrap_it_with_the_plastic_wrap_episode_18-train-bf_000800-af_001175-plus-0000277", "task": "Grab the croissant from the gray plate and wrap it with the plastic wrap", "image": [ "images/human_egodex/Grab_the_croissant_from_the_gray_plate_and_wrap_it_with_the_plastic_wrap/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Grab_the_croissant_from_the_gray_plate_and_wrap_it_with_the_plastic_wrap/episode_18/cam_high/frame_001241.jpg", "images/human_egodex/Grab_the_croissant_from_the_gray_plate_and_wrap_it_with_the_plastic_wrap/episode_18/cam_high/frame_000800.jpg", "images/human_egodex/Grab_the_croissant_from_the_gray_plate_and_wrap_it_with_the_plastic_wrap/episode_18/cam_high/frame_000800.jpg", "images/human_egodex/Grab_the_croissant_from_the_gray_plate_and_wrap_it_with_the_plastic_wrap/episode_18/cam_high/frame_000800.jpg", "images/human_egodex/Grab_the_croissant_from_the_gray_plate_and_wrap_it_with_the_plastic_wrap/episode_18/cam_high/frame_001175.jpg", "images/human_egodex/Grab_the_croissant_from_the_gray_plate_and_wrap_it_with_the_plastic_wrap/episode_18/cam_high/frame_001175.jpg", "images/human_egodex/Grab_the_croissant_from_the_gray_plate_and_wrap_it_with_the_plastic_wrap/episode_18/cam_high/frame_001175.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab the croissant from the gray plate and wrap it with the plastic wrap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab the croissant from the gray plate and wrap it with the plastic wrap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 88.2, "hop_absolute_value": 88.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background_episode_5-train-bf_000075-af_000175-plus-0000470", "task": "Disassemble the lobster puzzle on a metal table while sitting with a white background", "image": [ "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_5/cam_high/frame_000872.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_5/cam_high/frame_000075.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_5/cam_high/frame_000075.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_5/cam_high/frame_000075.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_5/cam_high/frame_000175.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_5/cam_high/frame_000175.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_5/cam_high/frame_000175.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble the lobster puzzle on a metal table while sitting with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble the lobster puzzle on a metal table while sitting with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.9, "hop_absolute_value": 12.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Insert_six_blocks_into_a_dump_container_on_a_blue_tablecloth_with_a_pink_background_episode_9-train-bf_000000-af_000150-plus-0000070", "task": "Insert six blocks into a dump container on a blue tablecloth with a pink background", "image": [ "images/human_egodex/Insert_six_blocks_into_a_dump_container_on_a_blue_tablecloth_with_a_pink_background/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Insert_six_blocks_into_a_dump_container_on_a_blue_tablecloth_with_a_pink_background/episode_9/cam_high/frame_000236.jpg", "images/human_egodex/Insert_six_blocks_into_a_dump_container_on_a_blue_tablecloth_with_a_pink_background/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Insert_six_blocks_into_a_dump_container_on_a_blue_tablecloth_with_a_pink_background/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Insert_six_blocks_into_a_dump_container_on_a_blue_tablecloth_with_a_pink_background/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Insert_six_blocks_into_a_dump_container_on_a_blue_tablecloth_with_a_pink_background/episode_9/cam_high/frame_000150.jpg", "images/human_egodex/Insert_six_blocks_into_a_dump_container_on_a_blue_tablecloth_with_a_pink_background/episode_9/cam_high/frame_000150.jpg", "images/human_egodex/Insert_six_blocks_into_a_dump_container_on_a_blue_tablecloth_with_a_pink_background/episode_9/cam_high/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert six blocks into a dump container on a blue tablecloth with a pink background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert six blocks into a dump container on a blue tablecloth with a pink background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Open_the_box_insert_the_plush_then_close_the_box_on_the_wooden_table_episode_0-train-bf_000050-af_000075-plus-0000058", "task": "Open the box insert the plush then close the box on the wooden table", "image": [ "images/human_egodex/Open_the_box_insert_the_plush_then_close_the_box_on_the_wooden_table/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Open_the_box_insert_the_plush_then_close_the_box_on_the_wooden_table/episode_0/cam_high/frame_000268.jpg", "images/human_egodex/Open_the_box_insert_the_plush_then_close_the_box_on_the_wooden_table/episode_0/cam_high/frame_000050.jpg", "images/human_egodex/Open_the_box_insert_the_plush_then_close_the_box_on_the_wooden_table/episode_0/cam_high/frame_000050.jpg", "images/human_egodex/Open_the_box_insert_the_plush_then_close_the_box_on_the_wooden_table/episode_0/cam_high/frame_000050.jpg", "images/human_egodex/Open_the_box_insert_the_plush_then_close_the_box_on_the_wooden_table/episode_0/cam_high/frame_000075.jpg", "images/human_egodex/Open_the_box_insert_the_plush_then_close_the_box_on_the_wooden_table/episode_0/cam_high/frame_000075.jpg", "images/human_egodex/Open_the_box_insert_the_plush_then_close_the_box_on_the_wooden_table/episode_0/cam_high/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the box insert the plush then close the box on the wooden table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the box insert the plush then close the box on the wooden table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_a_tennis_ball_from_a_container_while_sitting_at_a_table_with_a_white_background_episode_11-train-bf_000175-af_000275-plus-0000433", "task": "Remove a tennis ball from a container while sitting at a table with a white background", "image": [ "images/human_egodex/Remove_a_tennis_ball_from_a_container_while_sitting_at_a_table_with_a_white_background/episode_4/cam_high/frame_000000.jpg", "images/human_egodex/Remove_a_tennis_ball_from_a_container_while_sitting_at_a_table_with_a_white_background/episode_4/cam_high/frame_000835.jpg", "images/human_egodex/Remove_a_tennis_ball_from_a_container_while_sitting_at_a_table_with_a_white_background/episode_11/cam_high/frame_000175.jpg", "images/human_egodex/Remove_a_tennis_ball_from_a_container_while_sitting_at_a_table_with_a_white_background/episode_11/cam_high/frame_000175.jpg", "images/human_egodex/Remove_a_tennis_ball_from_a_container_while_sitting_at_a_table_with_a_white_background/episode_11/cam_high/frame_000175.jpg", "images/human_egodex/Remove_a_tennis_ball_from_a_container_while_sitting_at_a_table_with_a_white_background/episode_11/cam_high/frame_000275.jpg", "images/human_egodex/Remove_a_tennis_ball_from_a_container_while_sitting_at_a_table_with_a_white_background/episode_11/cam_high/frame_000275.jpg", "images/human_egodex/Remove_a_tennis_ball_from_a_container_while_sitting_at_a_table_with_a_white_background/episode_11/cam_high/frame_000275.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove a tennis ball from a container while sitting at a table with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove a tennis ball from a container while sitting at a table with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.5, "hop_absolute_value": 10.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unstack_ten_gray_cups_on_a_wooden_table_while_sitting_against_a_white_background_episode_1-train-bf_000250-af_000400-plus-0000186", "task": "Unstack ten gray cups on a wooden table while sitting against a white background", "image": [ "images/human_egodex/Unstack_ten_gray_cups_on_a_wooden_table_while_sitting_against_a_white_background/episode_1/cam_high/frame_000000.jpg", "images/human_egodex/Unstack_ten_gray_cups_on_a_wooden_table_while_sitting_against_a_white_background/episode_1/cam_high/frame_000475.jpg", "images/human_egodex/Unstack_ten_gray_cups_on_a_wooden_table_while_sitting_against_a_white_background/episode_1/cam_high/frame_000250.jpg", "images/human_egodex/Unstack_ten_gray_cups_on_a_wooden_table_while_sitting_against_a_white_background/episode_1/cam_high/frame_000250.jpg", "images/human_egodex/Unstack_ten_gray_cups_on_a_wooden_table_while_sitting_against_a_white_background/episode_1/cam_high/frame_000250.jpg", "images/human_egodex/Unstack_ten_gray_cups_on_a_wooden_table_while_sitting_against_a_white_background/episode_1/cam_high/frame_000400.jpg", "images/human_egodex/Unstack_ten_gray_cups_on_a_wooden_table_while_sitting_against_a_white_background/episode_1/cam_high/frame_000400.jpg", "images/human_egodex/Unstack_ten_gray_cups_on_a_wooden_table_while_sitting_against_a_white_background/episode_1/cam_high/frame_000400.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unstack ten gray cups on a wooden table while sitting against a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unstack ten gray cups on a wooden table while sitting against a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one_episode_11-train-bf_000100-af_000423-plus-0000214", "task": "Reverse the action by popping out the pressed bubbles on the square fidget Pop It toy one by one", "image": [ "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_3/cam_high/frame_000827.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_11/cam_high/frame_000100.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_11/cam_high/frame_000100.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_11/cam_high/frame_000100.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_11/cam_high/frame_000423.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_11/cam_high/frame_000423.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_11/cam_high/frame_000423.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Reverse the action by popping out the pressed bubbles on the square fidget Pop It toy one by one`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Reverse the action by popping out the pressed bubbles on the square fidget Pop It toy one by one`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unscrew_the_screw_from_the_fixture_using_fingers_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background_episode_3-train-bf_000075-af_000375-plus-0000213", "task": "Unscrew the screw from the fixture using fingers while sitting at a table with a pink tablecloth and lavender background", "image": [ "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_fingers_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_fingers_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000446.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_fingers_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000075.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_fingers_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000075.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_fingers_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000075.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_fingers_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000375.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_fingers_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000375.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_fingers_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unscrew the screw from the fixture using fingers while sitting at a table with a pink tablecloth and lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unscrew the screw from the fixture using fingers while sitting at a table with a pink tablecloth and lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting_episode_9-train-bf_000925-af_001045-plus-0000290", "task": "Disassemble tiles on a metal table with a lavender background while sitting", "image": [ "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_9/cam_high/frame_001045.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_9/cam_high/frame_000925.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_9/cam_high/frame_000925.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_9/cam_high/frame_000925.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_9/cam_high/frame_001045.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_9/cam_high/frame_001045.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_9/cam_high/frame_001045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble tiles on a metal table with a lavender background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble tiles on a metal table with a lavender background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_mini_piano_while_sitting_against_a_white_background_episode_15-train-bf_000450-af_000475-plus-0000420", "task": "Pick up the measuring tape and measure the dimensions of the mini piano while sitting against a white background", "image": [ "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_mini_piano_while_sitting_against_a_white_background/episode_15/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_mini_piano_while_sitting_against_a_white_background/episode_15/cam_high/frame_001001.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_mini_piano_while_sitting_against_a_white_background/episode_15/cam_high/frame_000450.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_mini_piano_while_sitting_against_a_white_background/episode_15/cam_high/frame_000450.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_mini_piano_while_sitting_against_a_white_background/episode_15/cam_high/frame_000450.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_mini_piano_while_sitting_against_a_white_background/episode_15/cam_high/frame_000475.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_mini_piano_while_sitting_against_a_white_background/episode_15/cam_high/frame_000475.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_mini_piano_while_sitting_against_a_white_background/episode_15/cam_high/frame_000475.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the measuring tape and measure the dimensions of the mini piano while sitting against a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the measuring tape and measure the dimensions of the mini piano while sitting against a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.5, "hop_absolute_value": 4.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unbraid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background_episode_14-train-bf_000125-af_000150-plus-0000189", "task": "Unbraid the three strands of ribbon while sitting against a brown background", "image": [ "images/human_egodex/Unbraid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_14/cam_high/frame_000000.jpg", "images/human_egodex/Unbraid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_14/cam_high/frame_000358.jpg", "images/human_egodex/Unbraid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_14/cam_high/frame_000125.jpg", "images/human_egodex/Unbraid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_14/cam_high/frame_000125.jpg", "images/human_egodex/Unbraid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_14/cam_high/frame_000125.jpg", "images/human_egodex/Unbraid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_14/cam_high/frame_000150.jpg", "images/human_egodex/Unbraid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_14/cam_high/frame_000150.jpg", "images/human_egodex/Unbraid_the_three_strands_of_ribbon_while_sitting_against_a_brown_background/episode_14/cam_high/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unbraid the three strands of ribbon while sitting against a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unbraid the three strands of ribbon while sitting against a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Set_up_the_board_game_by_placing_the_board_and_pieces_from_the_box_onto_the_wooden_table_with_a_red_background_episode_18-train-bf_000100-af_000250-plus-0000124", "task": "Set up the board game by placing the board and pieces from the box onto the wooden table with a red background", "image": [ "images/human_egodex/Set_up_the_board_game_by_placing_the_board_and_pieces_from_the_box_onto_the_wooden_table_with_a_red_background/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Set_up_the_board_game_by_placing_the_board_and_pieces_from_the_box_onto_the_wooden_table_with_a_red_background/episode_18/cam_high/frame_000302.jpg", "images/human_egodex/Set_up_the_board_game_by_placing_the_board_and_pieces_from_the_box_onto_the_wooden_table_with_a_red_background/episode_18/cam_high/frame_000100.jpg", "images/human_egodex/Set_up_the_board_game_by_placing_the_board_and_pieces_from_the_box_onto_the_wooden_table_with_a_red_background/episode_18/cam_high/frame_000100.jpg", "images/human_egodex/Set_up_the_board_game_by_placing_the_board_and_pieces_from_the_box_onto_the_wooden_table_with_a_red_background/episode_18/cam_high/frame_000100.jpg", "images/human_egodex/Set_up_the_board_game_by_placing_the_board_and_pieces_from_the_box_onto_the_wooden_table_with_a_red_background/episode_18/cam_high/frame_000250.jpg", "images/human_egodex/Set_up_the_board_game_by_placing_the_board_and_pieces_from_the_box_onto_the_wooden_table_with_a_red_background/episode_18/cam_high/frame_000250.jpg", "images/human_egodex/Set_up_the_board_game_by_placing_the_board_and_pieces_from_the_box_onto_the_wooden_table_with_a_red_background/episode_18/cam_high/frame_000250.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Set up the board game by placing the board and pieces from the box onto the wooden table with a red background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Set up the board game by placing the board and pieces from the box onto the wooden table with a red background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Braid_the_three_strands_of_hotpink_yarn_into_a_braid_while_sitting_against_a_brown_background_episode_0-train-bf_000325-af_000425-plus-0000469", "task": "Braid the three strands of hotpink yarn into a braid while sitting against a brown background", "image": [ "images/human_egodex/Braid_the_three_strands_of_hotpink_yarn_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Braid_the_three_strands_of_hotpink_yarn_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000763.jpg", "images/human_egodex/Braid_the_three_strands_of_hotpink_yarn_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000325.jpg", "images/human_egodex/Braid_the_three_strands_of_hotpink_yarn_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000325.jpg", "images/human_egodex/Braid_the_three_strands_of_hotpink_yarn_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000325.jpg", "images/human_egodex/Braid_the_three_strands_of_hotpink_yarn_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000425.jpg", "images/human_egodex/Braid_the_three_strands_of_hotpink_yarn_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000425.jpg", "images/human_egodex/Braid_the_three_strands_of_hotpink_yarn_into_a_braid_while_sitting_against_a_brown_background/episode_0/cam_high/frame_000425.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Braid the three strands of hotpink yarn into a braid while sitting against a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Braid the three strands of hotpink yarn into a braid while sitting against a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 23.5, "hop_absolute_value": 23.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_a_square_table_desk_made_of_wood_on_a_brown_background_while_sitting_episode_10-train-bf_000025-af_000100-plus-0000031", "task": "Disassemble a square table desk made of wood on a brown background while sitting", "image": [ "images/human_egodex/Disassemble_a_square_table_desk_made_of_wood_on_a_brown_background_while_sitting/episode_10/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_a_square_table_desk_made_of_wood_on_a_brown_background_while_sitting/episode_10/cam_high/frame_000161.jpg", "images/human_egodex/Disassemble_a_square_table_desk_made_of_wood_on_a_brown_background_while_sitting/episode_10/cam_high/frame_000025.jpg", "images/human_egodex/Disassemble_a_square_table_desk_made_of_wood_on_a_brown_background_while_sitting/episode_10/cam_high/frame_000025.jpg", "images/human_egodex/Disassemble_a_square_table_desk_made_of_wood_on_a_brown_background_while_sitting/episode_10/cam_high/frame_000025.jpg", "images/human_egodex/Disassemble_a_square_table_desk_made_of_wood_on_a_brown_background_while_sitting/episode_10/cam_high/frame_000100.jpg", "images/human_egodex/Disassemble_a_square_table_desk_made_of_wood_on_a_brown_background_while_sitting/episode_10/cam_high/frame_000100.jpg", "images/human_egodex/Disassemble_a_square_table_desk_made_of_wood_on_a_brown_background_while_sitting/episode_10/cam_high/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble a square table desk made of wood on a brown background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble a square table desk made of wood on a brown background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background_episode_12-train-bf_000125-af_000150-plus-0000338", "task": "Insert a USB into the hairbrush with the right hand while sitting against a blue background", "image": [ "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000000.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_001229.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000125.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000125.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000125.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000150.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000150.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert a USB into the hairbrush with the right hand while sitting against a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert a USB into the hairbrush with the right hand while sitting against a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.3, "hop_absolute_value": 2.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unwrap_a_croissant_from_plastic_wrap_episode_11-train-bf_000700-af_000798-plus-0000281", "task": "Unwrap a croissant from plastic wrap", "image": [ "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_11/cam_high/frame_000798.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_11/cam_high/frame_000700.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_11/cam_high/frame_000700.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_11/cam_high/frame_000700.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_11/cam_high/frame_000798.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_11/cam_high/frame_000798.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_11/cam_high/frame_000798.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unwrap a croissant from plastic wrap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unwrap a croissant from plastic wrap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_a_tennis_ball_from_among_three_tennis_balls_on_a_pink_tablecloth_while_sitting_against_a_lavender_background_episode_17-train-bf_000100-af_000175-plus-0000270", "task": "Remove a tennis ball from among three tennis balls on a pink tablecloth while sitting against a lavender background", "image": [ "images/human_egodex/Remove_a_tennis_ball_from_among_three_tennis_balls_on_a_pink_tablecloth_while_sitting_against_a_lavender_background/episode_17/cam_high/frame_000000.jpg", "images/human_egodex/Remove_a_tennis_ball_from_among_three_tennis_balls_on_a_pink_tablecloth_while_sitting_against_a_lavender_background/episode_17/cam_high/frame_000439.jpg", "images/human_egodex/Remove_a_tennis_ball_from_among_three_tennis_balls_on_a_pink_tablecloth_while_sitting_against_a_lavender_background/episode_17/cam_high/frame_000100.jpg", "images/human_egodex/Remove_a_tennis_ball_from_among_three_tennis_balls_on_a_pink_tablecloth_while_sitting_against_a_lavender_background/episode_17/cam_high/frame_000100.jpg", "images/human_egodex/Remove_a_tennis_ball_from_among_three_tennis_balls_on_a_pink_tablecloth_while_sitting_against_a_lavender_background/episode_17/cam_high/frame_000100.jpg", "images/human_egodex/Remove_a_tennis_ball_from_among_three_tennis_balls_on_a_pink_tablecloth_while_sitting_against_a_lavender_background/episode_17/cam_high/frame_000175.jpg", "images/human_egodex/Remove_a_tennis_ball_from_among_three_tennis_balls_on_a_pink_tablecloth_while_sitting_against_a_lavender_background/episode_17/cam_high/frame_000175.jpg", "images/human_egodex/Remove_a_tennis_ball_from_among_three_tennis_balls_on_a_pink_tablecloth_while_sitting_against_a_lavender_background/episode_17/cam_high/frame_000175.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove a tennis ball from among three tennis balls on a pink tablecloth while sitting against a lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove a tennis ball from among three tennis balls on a pink tablecloth while sitting against a lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 23.1, "hop_absolute_value": 23.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve_episode_18-train-bf_000925-af_000950-plus-0000396", "task": "Pick up a Pokemon or Vanguard card from the metal table and insert it into a clear binder sleeve", "image": [ "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_13/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_13/cam_high/frame_001740.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_18/cam_high/frame_000925.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_18/cam_high/frame_000925.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_18/cam_high/frame_000925.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_18/cam_high/frame_000950.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_18/cam_high/frame_000950.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_18/cam_high/frame_000950.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up a Pokemon or Vanguard card from the metal table and insert it into a clear binder sleeve`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up a Pokemon or Vanguard card from the metal table and insert it into a clear binder sleeve`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.3, "hop_absolute_value": 2.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_While_sitting_at_a_table_with_a_white_background_grab_a_sandwich_from_a_gray_plate_and_wrap_it_with_aluminum_foil_episode_18-train-bf_000075-af_000100-plus-0000377", "task": "While sitting at a table with a white background grab a sandwich from a gray plate and wrap it with aluminum foil", "image": [ "images/human_egodex/While_sitting_at_a_table_with_a_white_background_grab_a_sandwich_from_a_gray_plate_and_wrap_it_with_aluminum_foil/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/While_sitting_at_a_table_with_a_white_background_grab_a_sandwich_from_a_gray_plate_and_wrap_it_with_aluminum_foil/episode_7/cam_high/frame_001047.jpg", "images/human_egodex/While_sitting_at_a_table_with_a_white_background_grab_a_sandwich_from_a_gray_plate_and_wrap_it_with_aluminum_foil/episode_18/cam_high/frame_000075.jpg", "images/human_egodex/While_sitting_at_a_table_with_a_white_background_grab_a_sandwich_from_a_gray_plate_and_wrap_it_with_aluminum_foil/episode_18/cam_high/frame_000075.jpg", "images/human_egodex/While_sitting_at_a_table_with_a_white_background_grab_a_sandwich_from_a_gray_plate_and_wrap_it_with_aluminum_foil/episode_18/cam_high/frame_000075.jpg", "images/human_egodex/While_sitting_at_a_table_with_a_white_background_grab_a_sandwich_from_a_gray_plate_and_wrap_it_with_aluminum_foil/episode_18/cam_high/frame_000100.jpg", "images/human_egodex/While_sitting_at_a_table_with_a_white_background_grab_a_sandwich_from_a_gray_plate_and_wrap_it_with_aluminum_foil/episode_18/cam_high/frame_000100.jpg", "images/human_egodex/While_sitting_at_a_table_with_a_white_background_grab_a_sandwich_from_a_gray_plate_and_wrap_it_with_aluminum_foil/episode_18/cam_high/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`While sitting at a table with a white background grab a sandwich from a gray plate and wrap it with aluminum foil`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `While sitting at a table with a white background grab a sandwich from a gray plate and wrap it with aluminum foil`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.6, "hop_absolute_value": 1.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Boil_water_in_pot_take_egg_from_carton_boil_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink_episode_9-train-bf_000050-af_000150-plus-0000442", "task": "Boil water in pot take egg from carton boil egg serve egg in gray bowl on round table and place pot and ladle in sink", "image": [ "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_18/cam_high/frame_001376.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_9/cam_high/frame_000050.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_9/cam_high/frame_000050.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_9/cam_high/frame_000050.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_9/cam_high/frame_000150.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_9/cam_high/frame_000150.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_9/cam_high/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Boil water in pot take egg from carton boil egg serve egg in gray bowl on round table and place pot and ladle in sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Boil water in pot take egg from carton boil egg serve egg in gray bowl on round table and place pot and ladle in sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.3, "hop_absolute_value": 10.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_clear_tupperware_from_the_table_with_the_right_hand_and_stack_it_on_the_shelf_episode_14-train-bf_000400-af_000425-plus-0000138", "task": "Pick up the clear tupperware from the table with the right hand and stack it on the shelf", "image": [ "images/human_egodex/Pick_up_the_clear_tupperware_from_the_table_with_the_right_hand_and_stack_it_on_the_shelf/episode_14/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_clear_tupperware_from_the_table_with_the_right_hand_and_stack_it_on_the_shelf/episode_14/cam_high/frame_000522.jpg", "images/human_egodex/Pick_up_the_clear_tupperware_from_the_table_with_the_right_hand_and_stack_it_on_the_shelf/episode_14/cam_high/frame_000400.jpg", "images/human_egodex/Pick_up_the_clear_tupperware_from_the_table_with_the_right_hand_and_stack_it_on_the_shelf/episode_14/cam_high/frame_000400.jpg", "images/human_egodex/Pick_up_the_clear_tupperware_from_the_table_with_the_right_hand_and_stack_it_on_the_shelf/episode_14/cam_high/frame_000400.jpg", "images/human_egodex/Pick_up_the_clear_tupperware_from_the_table_with_the_right_hand_and_stack_it_on_the_shelf/episode_14/cam_high/frame_000425.jpg", "images/human_egodex/Pick_up_the_clear_tupperware_from_the_table_with_the_right_hand_and_stack_it_on_the_shelf/episode_14/cam_high/frame_000425.jpg", "images/human_egodex/Pick_up_the_clear_tupperware_from_the_table_with_the_right_hand_and_stack_it_on_the_shelf/episode_14/cam_high/frame_000425.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the clear tupperware from the table with the right hand and stack it on the shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the clear tupperware from the table with the right hand and stack it on the shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_lids_from_three_cups_on_a_wooden_table_while_sitting_with_a_brown_background_episode_18-train-bf_000025-af_000225-plus-0000111", "task": "Remove lids from three cups on a wooden table while sitting with a brown background", "image": [ "images/human_egodex/Remove_lids_from_three_cups_on_a_wooden_table_while_sitting_with_a_brown_background/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Remove_lids_from_three_cups_on_a_wooden_table_while_sitting_with_a_brown_background/episode_18/cam_high/frame_000282.jpg", "images/human_egodex/Remove_lids_from_three_cups_on_a_wooden_table_while_sitting_with_a_brown_background/episode_18/cam_high/frame_000025.jpg", "images/human_egodex/Remove_lids_from_three_cups_on_a_wooden_table_while_sitting_with_a_brown_background/episode_18/cam_high/frame_000025.jpg", "images/human_egodex/Remove_lids_from_three_cups_on_a_wooden_table_while_sitting_with_a_brown_background/episode_18/cam_high/frame_000025.jpg", "images/human_egodex/Remove_lids_from_three_cups_on_a_wooden_table_while_sitting_with_a_brown_background/episode_18/cam_high/frame_000225.jpg", "images/human_egodex/Remove_lids_from_three_cups_on_a_wooden_table_while_sitting_with_a_brown_background/episode_18/cam_high/frame_000225.jpg", "images/human_egodex/Remove_lids_from_three_cups_on_a_wooden_table_while_sitting_with_a_brown_background/episode_18/cam_high/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove lids from three cups on a wooden table while sitting with a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove lids from three cups on a wooden table while sitting with a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Assemble_six_legos_on_a_blue_tablecloth_while_sitting_with_a_pink_background_episode_8-train-bf_000100-af_000154-plus-0000036", "task": "Assemble six legos on a blue tablecloth while sitting with a pink background", "image": [ "images/human_egodex/Assemble_six_legos_on_a_blue_tablecloth_while_sitting_with_a_pink_background/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Assemble_six_legos_on_a_blue_tablecloth_while_sitting_with_a_pink_background/episode_8/cam_high/frame_000154.jpg", "images/human_egodex/Assemble_six_legos_on_a_blue_tablecloth_while_sitting_with_a_pink_background/episode_8/cam_high/frame_000100.jpg", "images/human_egodex/Assemble_six_legos_on_a_blue_tablecloth_while_sitting_with_a_pink_background/episode_8/cam_high/frame_000100.jpg", "images/human_egodex/Assemble_six_legos_on_a_blue_tablecloth_while_sitting_with_a_pink_background/episode_8/cam_high/frame_000100.jpg", "images/human_egodex/Assemble_six_legos_on_a_blue_tablecloth_while_sitting_with_a_pink_background/episode_8/cam_high/frame_000154.jpg", "images/human_egodex/Assemble_six_legos_on_a_blue_tablecloth_while_sitting_with_a_pink_background/episode_8/cam_high/frame_000154.jpg", "images/human_egodex/Assemble_six_legos_on_a_blue_tablecloth_while_sitting_with_a_pink_background/episode_8/cam_high/frame_000154.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Assemble six legos on a blue tablecloth while sitting with a pink background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Assemble six legos on a blue tablecloth while sitting with a pink background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Braid_the_three_strands_of_salmonpink_yarn_episode_6-train-bf_000175-af_000200-plus-0000217", "task": "Braid the three strands of salmonpink yarn", "image": [ "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_6/cam_high/frame_000394.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_6/cam_high/frame_000175.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_6/cam_high/frame_000175.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_6/cam_high/frame_000175.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_6/cam_high/frame_000200.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_6/cam_high/frame_000200.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_6/cam_high/frame_000200.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Braid the three strands of salmonpink yarn`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Braid the three strands of salmonpink yarn`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Take_the_crayon_out_of_the_box_color_on_the_coloring_page_and_put_the_crayon_on_the_table_episode_2-train-bf_000075-af_000175-plus-0000091", "task": "Take the crayon out of the box color on the coloring page and put the crayon on the table", "image": [ "images/human_egodex/Take_the_crayon_out_of_the_box_color_on_the_coloring_page_and_put_the_crayon_on_the_table/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Take_the_crayon_out_of_the_box_color_on_the_coloring_page_and_put_the_crayon_on_the_table/episode_2/cam_high/frame_000299.jpg", "images/human_egodex/Take_the_crayon_out_of_the_box_color_on_the_coloring_page_and_put_the_crayon_on_the_table/episode_2/cam_high/frame_000075.jpg", "images/human_egodex/Take_the_crayon_out_of_the_box_color_on_the_coloring_page_and_put_the_crayon_on_the_table/episode_2/cam_high/frame_000075.jpg", "images/human_egodex/Take_the_crayon_out_of_the_box_color_on_the_coloring_page_and_put_the_crayon_on_the_table/episode_2/cam_high/frame_000075.jpg", "images/human_egodex/Take_the_crayon_out_of_the_box_color_on_the_coloring_page_and_put_the_crayon_on_the_table/episode_2/cam_high/frame_000175.jpg", "images/human_egodex/Take_the_crayon_out_of_the_box_color_on_the_coloring_page_and_put_the_crayon_on_the_table/episode_2/cam_high/frame_000175.jpg", "images/human_egodex/Take_the_crayon_out_of_the_box_color_on_the_coloring_page_and_put_the_crayon_on_the_table/episode_2/cam_high/frame_000175.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the crayon out of the box color on the coloring page and put the crayon on the table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the crayon out of the box color on the coloring page and put the crayon on the table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background_episode_6-train-bf_000225-af_000475-plus-0000217", "task": "Screw the screw into the fixture using an allen key while sitting at a table with a blue tablecloth and pink background", "image": [ "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_17/cam_high/frame_000000.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_17/cam_high/frame_000582.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_6/cam_high/frame_000225.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_6/cam_high/frame_000225.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_6/cam_high/frame_000225.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_6/cam_high/frame_000475.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_6/cam_high/frame_000475.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_6/cam_high/frame_000475.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Screw the screw into the fixture using an allen key while sitting at a table with a blue tablecloth and pink background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Screw the screw into the fixture using an allen key while sitting at a table with a blue tablecloth and pink background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background_episode_0-train-bf_000125-af_000150-plus-0000328", "task": "Screw the screw into the fixture using an allen key while sitting at a table with a pink tablecloth and lavender background", "image": [ "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_0/cam_high/frame_001139.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_0/cam_high/frame_000125.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_0/cam_high/frame_000125.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_0/cam_high/frame_000125.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_0/cam_high/frame_000150.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_0/cam_high/frame_000150.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_using_an_allen_key_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_0/cam_high/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Screw the screw into the fixture using an allen key while sitting at a table with a pink tablecloth and lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Screw the screw into the fixture using an allen key while sitting at a table with a pink tablecloth and lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.5, "hop_absolute_value": 2.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_a_desk_while_sitting_at_a_table_with_a_green_tablecloth_and_white_background_episode_14-train-bf_000000-af_000050-plus-0000018", "task": "Disassemble a desk while sitting at a table with a green tablecloth and white background", "image": [ "images/human_egodex/Disassemble_a_desk_while_sitting_at_a_table_with_a_green_tablecloth_and_white_background/episode_14/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_a_desk_while_sitting_at_a_table_with_a_green_tablecloth_and_white_background/episode_14/cam_high/frame_000141.jpg", "images/human_egodex/Disassemble_a_desk_while_sitting_at_a_table_with_a_green_tablecloth_and_white_background/episode_14/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_a_desk_while_sitting_at_a_table_with_a_green_tablecloth_and_white_background/episode_14/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_a_desk_while_sitting_at_a_table_with_a_green_tablecloth_and_white_background/episode_14/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_a_desk_while_sitting_at_a_table_with_a_green_tablecloth_and_white_background/episode_14/cam_high/frame_000050.jpg", "images/human_egodex/Disassemble_a_desk_while_sitting_at_a_table_with_a_green_tablecloth_and_white_background/episode_14/cam_high/frame_000050.jpg", "images/human_egodex/Disassemble_a_desk_while_sitting_at_a_table_with_a_green_tablecloth_and_white_background/episode_14/cam_high/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble a desk while sitting at a table with a green tablecloth and white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble a desk while sitting at a table with a green tablecloth and white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_the_puzzle_episode_5-train-bf_000450-af_000550-plus-0000189", "task": "Pick up the puzzle pieces from a wooden table with the right hand and assemble the puzzle", "image": [ "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000642.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000450.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000450.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000450.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000550.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000550.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000550.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the puzzle pieces from a wooden table with the right hand and assemble the puzzle`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the puzzle pieces from a wooden table with the right hand and assemble the puzzle`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.1, "hop_absolute_value": 57.1, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table_episode_14-train-bf_000100-af_000125-plus-0000333", "task": "Remove the Pokemon card from the black card sleeve and place it back on the metal table", "image": [ "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_14/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_14/cam_high/frame_001120.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_14/cam_high/frame_000100.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_14/cam_high/frame_000100.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_14/cam_high/frame_000100.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_14/cam_high/frame_000125.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_14/cam_high/frame_000125.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_14/cam_high/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the Pokemon card from the black card sleeve and place it back on the metal table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the Pokemon card from the black card sleeve and place it back on the metal table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.5, "hop_absolute_value": 2.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unfold_the_piece_of_paper_while_sitting_at_a_pink_tablecloth_with_a_lavender_background_episode_16-train-bf_000075-af_000100-plus-0000316", "task": "Unfold the piece of paper while sitting at a pink tablecloth with a lavender background", "image": [ "images/human_egodex/Unfold_the_piece_of_paper_while_sitting_at_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000000.jpg", "images/human_egodex/Unfold_the_piece_of_paper_while_sitting_at_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000718.jpg", "images/human_egodex/Unfold_the_piece_of_paper_while_sitting_at_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000075.jpg", "images/human_egodex/Unfold_the_piece_of_paper_while_sitting_at_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000075.jpg", "images/human_egodex/Unfold_the_piece_of_paper_while_sitting_at_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000075.jpg", "images/human_egodex/Unfold_the_piece_of_paper_while_sitting_at_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000100.jpg", "images/human_egodex/Unfold_the_piece_of_paper_while_sitting_at_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000100.jpg", "images/human_egodex/Unfold_the_piece_of_paper_while_sitting_at_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the piece of paper while sitting at a pink tablecloth with a lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the piece of paper while sitting at a pink tablecloth with a lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.0, "hop_absolute_value": 4.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Assemble_soft_legos_into_any_sort_of_tower_on_a_white_background_while_sitting_episode_15-train-bf_000750-af_000775-plus-0000404", "task": "Assemble soft legos into any sort of tower on a white background while sitting", "image": [ "images/human_egodex/Assemble_soft_legos_into_any_sort_of_tower_on_a_white_background_while_sitting/episode_10/cam_high/frame_000000.jpg", "images/human_egodex/Assemble_soft_legos_into_any_sort_of_tower_on_a_white_background_while_sitting/episode_10/cam_high/frame_002436.jpg", "images/human_egodex/Assemble_soft_legos_into_any_sort_of_tower_on_a_white_background_while_sitting/episode_15/cam_high/frame_000750.jpg", "images/human_egodex/Assemble_soft_legos_into_any_sort_of_tower_on_a_white_background_while_sitting/episode_15/cam_high/frame_000750.jpg", "images/human_egodex/Assemble_soft_legos_into_any_sort_of_tower_on_a_white_background_while_sitting/episode_15/cam_high/frame_000750.jpg", "images/human_egodex/Assemble_soft_legos_into_any_sort_of_tower_on_a_white_background_while_sitting/episode_15/cam_high/frame_000775.jpg", "images/human_egodex/Assemble_soft_legos_into_any_sort_of_tower_on_a_white_background_while_sitting/episode_15/cam_high/frame_000775.jpg", "images/human_egodex/Assemble_soft_legos_into_any_sort_of_tower_on_a_white_background_while_sitting/episode_15/cam_high/frame_000775.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Assemble soft legos into any sort of tower on a white background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Assemble soft legos into any sort of tower on a white background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 0.8, "hop_absolute_value": 0.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unwrap_a_croissant_from_plastic_wrap_episode_8-train-bf_000150-af_000175-plus-0000137", "task": "Unwrap a croissant from plastic wrap", "image": [ "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_8/cam_high/frame_000588.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_8/cam_high/frame_000150.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_8/cam_high/frame_000150.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_8/cam_high/frame_000150.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_8/cam_high/frame_000175.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_8/cam_high/frame_000175.jpg", "images/human_egodex/Unwrap_a_croissant_from_plastic_wrap/episode_8/cam_high/frame_000175.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unwrap a croissant from plastic wrap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unwrap a croissant from plastic wrap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.9, "hop_absolute_value": 5.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_structures_on_a_wooden_table_while_sitting_against_a_white_background_episode_7-train-bf_000275-af_000300-plus-0000349", "task": "Disassemble structures on a wooden table while sitting against a white background", "image": [ "images/human_egodex/Disassemble_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_2/cam_high/frame_000617.jpg", "images/human_egodex/Disassemble_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_7/cam_high/frame_000275.jpg", "images/human_egodex/Disassemble_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_7/cam_high/frame_000275.jpg", "images/human_egodex/Disassemble_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_7/cam_high/frame_000275.jpg", "images/human_egodex/Disassemble_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_7/cam_high/frame_000300.jpg", "images/human_egodex/Disassemble_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_7/cam_high/frame_000300.jpg", "images/human_egodex/Disassemble_structures_on_a_wooden_table_while_sitting_against_a_white_background/episode_7/cam_high/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble structures on a wooden table while sitting against a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble structures on a wooden table while sitting against a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.0, "hop_absolute_value": 5.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background_episode_10-train-bf_000575-af_000600-plus-0000390", "task": "Remove the USB from the hairbrush using the left hand while sitting against a blue background", "image": [ "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_10/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_10/cam_high/frame_000745.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_10/cam_high/frame_000575.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_10/cam_high/frame_000575.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_10/cam_high/frame_000575.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_10/cam_high/frame_000600.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_10/cam_high/frame_000600.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_left_hand_while_sitting_against_a_blue_background/episode_10/cam_high/frame_000600.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the USB from the hairbrush using the left hand while sitting against a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the USB from the hairbrush using the left hand while sitting against a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Fold_a_medium_sized_tshirt_on_a_green_tablecloth_while_sitting_episode_2-train-bf_000425-af_000450-plus-0000331", "task": "Fold a medium sized tshirt on a green tablecloth while sitting", "image": [ "images/human_egodex/Fold_a_medium_sized_tshirt_on_a_green_tablecloth_while_sitting/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Fold_a_medium_sized_tshirt_on_a_green_tablecloth_while_sitting/episode_2/cam_high/frame_000528.jpg", "images/human_egodex/Fold_a_medium_sized_tshirt_on_a_green_tablecloth_while_sitting/episode_2/cam_high/frame_000425.jpg", "images/human_egodex/Fold_a_medium_sized_tshirt_on_a_green_tablecloth_while_sitting/episode_2/cam_high/frame_000425.jpg", "images/human_egodex/Fold_a_medium_sized_tshirt_on_a_green_tablecloth_while_sitting/episode_2/cam_high/frame_000425.jpg", "images/human_egodex/Fold_a_medium_sized_tshirt_on_a_green_tablecloth_while_sitting/episode_2/cam_high/frame_000450.jpg", "images/human_egodex/Fold_a_medium_sized_tshirt_on_a_green_tablecloth_while_sitting/episode_2/cam_high/frame_000450.jpg", "images/human_egodex/Fold_a_medium_sized_tshirt_on_a_green_tablecloth_while_sitting/episode_2/cam_high/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold a medium sized tshirt on a green tablecloth while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold a medium sized tshirt on a green tablecloth while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background_episode_12-train-bf_000200-af_000225-plus-0000077", "task": "Disassemble eighteen tiles on a wooden table while sitting against a white background", "image": [ "images/human_egodex/Disassemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_12/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_12/cam_high/frame_000299.jpg", "images/human_egodex/Disassemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_12/cam_high/frame_000200.jpg", "images/human_egodex/Disassemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_12/cam_high/frame_000200.jpg", "images/human_egodex/Disassemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_12/cam_high/frame_000200.jpg", "images/human_egodex/Disassemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_12/cam_high/frame_000225.jpg", "images/human_egodex/Disassemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_12/cam_high/frame_000225.jpg", "images/human_egodex/Disassemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_12/cam_high/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble eighteen tiles on a wooden table while sitting against a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble eighteen tiles on a wooden table while sitting against a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unscrew_the_screw_from_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background_episode_3-train-bf_000175-af_000225-plus-0000129", "task": "Unscrew the screw from the fixture while sitting at a table with a blue tablecloth and pink background", "image": [ "images/human_egodex/Unscrew_the_screw_from_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_3/cam_high/frame_000472.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_3/cam_high/frame_000175.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_3/cam_high/frame_000175.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_3/cam_high/frame_000175.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_3/cam_high/frame_000225.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_3/cam_high/frame_000225.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_3/cam_high/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unscrew the screw from the fixture while sitting at a table with a blue tablecloth and pink background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unscrew the screw from the fixture while sitting at a table with a blue tablecloth and pink background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 18.2, "hop_absolute_value": 18.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_toy_stove_door_episode_5-train-bf_000225-af_000300-plus-0000400", "task": "Pick up the measuring tape and measure the dimensions of the toy stove door", "image": [ "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_toy_stove_door/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_toy_stove_door/episode_9/cam_high/frame_000542.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_toy_stove_door/episode_5/cam_high/frame_000225.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_toy_stove_door/episode_5/cam_high/frame_000225.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_toy_stove_door/episode_5/cam_high/frame_000225.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_toy_stove_door/episode_5/cam_high/frame_000300.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_toy_stove_door/episode_5/cam_high/frame_000300.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_toy_stove_door/episode_5/cam_high/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the measuring tape and measure the dimensions of the toy stove door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the measuring tape and measure the dimensions of the toy stove door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.4, "hop_absolute_value": 9.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Use_chopsticks_to_move_strings_from_one_bowl_to_another_while_sitting_at_a_table_with_a_green_tablecloth_episode_3-train-bf_000650-af_000700-plus-0000416", "task": "Use chopsticks to move strings from one bowl to another while sitting at a table with a green tablecloth", "image": [ "images/human_egodex/Use_chopsticks_to_move_strings_from_one_bowl_to_another_while_sitting_at_a_table_with_a_green_tablecloth/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Use_chopsticks_to_move_strings_from_one_bowl_to_another_while_sitting_at_a_table_with_a_green_tablecloth/episode_3/cam_high/frame_002448.jpg", "images/human_egodex/Use_chopsticks_to_move_strings_from_one_bowl_to_another_while_sitting_at_a_table_with_a_green_tablecloth/episode_3/cam_high/frame_000650.jpg", "images/human_egodex/Use_chopsticks_to_move_strings_from_one_bowl_to_another_while_sitting_at_a_table_with_a_green_tablecloth/episode_3/cam_high/frame_000650.jpg", "images/human_egodex/Use_chopsticks_to_move_strings_from_one_bowl_to_another_while_sitting_at_a_table_with_a_green_tablecloth/episode_3/cam_high/frame_000650.jpg", "images/human_egodex/Use_chopsticks_to_move_strings_from_one_bowl_to_another_while_sitting_at_a_table_with_a_green_tablecloth/episode_3/cam_high/frame_000700.jpg", "images/human_egodex/Use_chopsticks_to_move_strings_from_one_bowl_to_another_while_sitting_at_a_table_with_a_green_tablecloth/episode_3/cam_high/frame_000700.jpg", "images/human_egodex/Use_chopsticks_to_move_strings_from_one_bowl_to_another_while_sitting_at_a_table_with_a_green_tablecloth/episode_3/cam_high/frame_000700.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use chopsticks to move strings from one bowl to another while sitting at a table with a green tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use chopsticks to move strings from one bowl to another while sitting at a table with a green tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.8, "hop_absolute_value": 2.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Stretch_a_short_rubber_band_and_wrap_it_around_a_group_of_pens_while_sitting_at_a_table_with_a_lavender_background_episode_11-train-bf_000075-af_000100-plus-0000031", "task": "Stretch a short rubber band and wrap it around a group of pens while sitting at a table with a lavender background", "image": [ "images/human_egodex/Stretch_a_short_rubber_band_and_wrap_it_around_a_group_of_pens_while_sitting_at_a_table_with_a_lavender_background/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Stretch_a_short_rubber_band_and_wrap_it_around_a_group_of_pens_while_sitting_at_a_table_with_a_lavender_background/episode_0/cam_high/frame_000252.jpg", "images/human_egodex/Stretch_a_short_rubber_band_and_wrap_it_around_a_group_of_pens_while_sitting_at_a_table_with_a_lavender_background/episode_11/cam_high/frame_000075.jpg", "images/human_egodex/Stretch_a_short_rubber_band_and_wrap_it_around_a_group_of_pens_while_sitting_at_a_table_with_a_lavender_background/episode_11/cam_high/frame_000075.jpg", "images/human_egodex/Stretch_a_short_rubber_band_and_wrap_it_around_a_group_of_pens_while_sitting_at_a_table_with_a_lavender_background/episode_11/cam_high/frame_000075.jpg", "images/human_egodex/Stretch_a_short_rubber_band_and_wrap_it_around_a_group_of_pens_while_sitting_at_a_table_with_a_lavender_background/episode_11/cam_high/frame_000100.jpg", "images/human_egodex/Stretch_a_short_rubber_band_and_wrap_it_around_a_group_of_pens_while_sitting_at_a_table_with_a_lavender_background/episode_11/cam_high/frame_000100.jpg", "images/human_egodex/Stretch_a_short_rubber_band_and_wrap_it_around_a_group_of_pens_while_sitting_at_a_table_with_a_lavender_background/episode_11/cam_high/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Stretch a short rubber band and wrap it around a group of pens while sitting at a table with a lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Stretch a short rubber band and wrap it around a group of pens while sitting at a table with a lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_a_square_table_from_furniture_bench_while_sitting_at_a_green_tablecloth_covered_table_with_a_white_background_episode_11-train-bf_000050-af_000100-plus-0000063", "task": "Disassemble a square table from furniture bench while sitting at a green tablecloth covered table with a white background", "image": [ "images/human_egodex/Disassemble_a_square_table_from_furniture_bench_while_sitting_at_a_green_tablecloth_covered_table_with_a_white_background/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_a_square_table_from_furniture_bench_while_sitting_at_a_green_tablecloth_covered_table_with_a_white_background/episode_11/cam_high/frame_000256.jpg", "images/human_egodex/Disassemble_a_square_table_from_furniture_bench_while_sitting_at_a_green_tablecloth_covered_table_with_a_white_background/episode_11/cam_high/frame_000050.jpg", "images/human_egodex/Disassemble_a_square_table_from_furniture_bench_while_sitting_at_a_green_tablecloth_covered_table_with_a_white_background/episode_11/cam_high/frame_000050.jpg", "images/human_egodex/Disassemble_a_square_table_from_furniture_bench_while_sitting_at_a_green_tablecloth_covered_table_with_a_white_background/episode_11/cam_high/frame_000050.jpg", "images/human_egodex/Disassemble_a_square_table_from_furniture_bench_while_sitting_at_a_green_tablecloth_covered_table_with_a_white_background/episode_11/cam_high/frame_000100.jpg", "images/human_egodex/Disassemble_a_square_table_from_furniture_bench_while_sitting_at_a_green_tablecloth_covered_table_with_a_white_background/episode_11/cam_high/frame_000100.jpg", "images/human_egodex/Disassemble_a_square_table_from_furniture_bench_while_sitting_at_a_green_tablecloth_covered_table_with_a_white_background/episode_11/cam_high/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble a square table from furniture bench while sitting at a green tablecloth covered table with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble a square table from furniture bench while sitting at a green tablecloth covered table with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_own_USB_from_black_USB_while_sitting_at_a_table_with_a_green_tablecloth_and_blue_background_episode_11-train-bf_000025-af_000050-plus-0000055", "task": "Remove own USB from black USB while sitting at a table with a green tablecloth and blue background", "image": [ "images/human_egodex/Remove_own_USB_from_black_USB_while_sitting_at_a_table_with_a_green_tablecloth_and_blue_background/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Remove_own_USB_from_black_USB_while_sitting_at_a_table_with_a_green_tablecloth_and_blue_background/episode_5/cam_high/frame_000239.jpg", "images/human_egodex/Remove_own_USB_from_black_USB_while_sitting_at_a_table_with_a_green_tablecloth_and_blue_background/episode_11/cam_high/frame_000025.jpg", "images/human_egodex/Remove_own_USB_from_black_USB_while_sitting_at_a_table_with_a_green_tablecloth_and_blue_background/episode_11/cam_high/frame_000025.jpg", "images/human_egodex/Remove_own_USB_from_black_USB_while_sitting_at_a_table_with_a_green_tablecloth_and_blue_background/episode_11/cam_high/frame_000025.jpg", "images/human_egodex/Remove_own_USB_from_black_USB_while_sitting_at_a_table_with_a_green_tablecloth_and_blue_background/episode_11/cam_high/frame_000050.jpg", "images/human_egodex/Remove_own_USB_from_black_USB_while_sitting_at_a_table_with_a_green_tablecloth_and_blue_background/episode_11/cam_high/frame_000050.jpg", "images/human_egodex/Remove_own_USB_from_black_USB_while_sitting_at_a_table_with_a_green_tablecloth_and_blue_background/episode_11/cam_high/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove own USB from black USB while sitting at a table with a green tablecloth and blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove own USB from black USB while sitting at a table with a green tablecloth and blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_crayon_from_the_table_place_it_back_into_the_box_and_stop_coloring_episode_10-train-bf_000200-af_000225-plus-0000078", "task": "Pick up the crayon from the table place it back into the box and stop coloring", "image": [ "images/human_egodex/Pick_up_the_crayon_from_the_table_place_it_back_into_the_box_and_stop_coloring/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_crayon_from_the_table_place_it_back_into_the_box_and_stop_coloring/episode_9/cam_high/frame_000299.jpg", "images/human_egodex/Pick_up_the_crayon_from_the_table_place_it_back_into_the_box_and_stop_coloring/episode_10/cam_high/frame_000200.jpg", "images/human_egodex/Pick_up_the_crayon_from_the_table_place_it_back_into_the_box_and_stop_coloring/episode_10/cam_high/frame_000200.jpg", "images/human_egodex/Pick_up_the_crayon_from_the_table_place_it_back_into_the_box_and_stop_coloring/episode_10/cam_high/frame_000200.jpg", "images/human_egodex/Pick_up_the_crayon_from_the_table_place_it_back_into_the_box_and_stop_coloring/episode_10/cam_high/frame_000225.jpg", "images/human_egodex/Pick_up_the_crayon_from_the_table_place_it_back_into_the_box_and_stop_coloring/episode_10/cam_high/frame_000225.jpg", "images/human_egodex/Pick_up_the_crayon_from_the_table_place_it_back_into_the_box_and_stop_coloring/episode_10/cam_high/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the crayon from the table place it back into the box and stop coloring`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the crayon from the table place it back into the box and stop coloring`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unstake_the_lego_bricks_to_disassemble_the_built_form_on_a_blue_tablecloth_episode_11-train-bf_000275-af_000300-plus-0000224", "task": "Unstake the lego bricks to disassemble the built form on a blue tablecloth", "image": [ "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_built_form_on_a_blue_tablecloth/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_built_form_on_a_blue_tablecloth/episode_11/cam_high/frame_000398.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_built_form_on_a_blue_tablecloth/episode_11/cam_high/frame_000275.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_built_form_on_a_blue_tablecloth/episode_11/cam_high/frame_000275.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_built_form_on_a_blue_tablecloth/episode_11/cam_high/frame_000275.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_built_form_on_a_blue_tablecloth/episode_11/cam_high/frame_000300.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_built_form_on_a_blue_tablecloth/episode_11/cam_high/frame_000300.jpg", "images/human_egodex/Unstake_the_lego_bricks_to_disassemble_the_built_form_on_a_blue_tablecloth/episode_11/cam_high/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unstake the lego bricks to disassemble the built form on a blue tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unstake the lego bricks to disassemble the built form on a blue tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_a_black_and_a_white_reversi_disc_and_place_them_on_the_reversi_board_which_is_on_a_metal_table_episode_14-train-bf_000100-af_000175-plus-0000083", "task": "Pick up a black and a white reversi disc and place them on the reversi board which is on a metal table", "image": [ "images/human_egodex/Pick_up_a_black_and_a_white_reversi_disc_and_place_them_on_the_reversi_board_which_is_on_a_metal_table/episode_19/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_a_black_and_a_white_reversi_disc_and_place_them_on_the_reversi_board_which_is_on_a_metal_table/episode_19/cam_high/frame_000299.jpg", "images/human_egodex/Pick_up_a_black_and_a_white_reversi_disc_and_place_them_on_the_reversi_board_which_is_on_a_metal_table/episode_14/cam_high/frame_000100.jpg", "images/human_egodex/Pick_up_a_black_and_a_white_reversi_disc_and_place_them_on_the_reversi_board_which_is_on_a_metal_table/episode_14/cam_high/frame_000100.jpg", "images/human_egodex/Pick_up_a_black_and_a_white_reversi_disc_and_place_them_on_the_reversi_board_which_is_on_a_metal_table/episode_14/cam_high/frame_000100.jpg", "images/human_egodex/Pick_up_a_black_and_a_white_reversi_disc_and_place_them_on_the_reversi_board_which_is_on_a_metal_table/episode_14/cam_high/frame_000175.jpg", "images/human_egodex/Pick_up_a_black_and_a_white_reversi_disc_and_place_them_on_the_reversi_board_which_is_on_a_metal_table/episode_14/cam_high/frame_000175.jpg", "images/human_egodex/Pick_up_a_black_and_a_white_reversi_disc_and_place_them_on_the_reversi_board_which_is_on_a_metal_table/episode_14/cam_high/frame_000175.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up a black and a white reversi disc and place them on the reversi board which is on a metal table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up a black and a white reversi disc and place them on the reversi board which is on a metal table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 42.9, "hop_absolute_value": 42.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_laptop_box_while_sitting_against_a_white_background_episode_19-train-bf_000025-af_000050-plus-0000147", "task": "Pick up the measuring tape and measure the dimensions of the laptop box while sitting against a white background", "image": [ "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_laptop_box_while_sitting_against_a_white_background/episode_19/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_laptop_box_while_sitting_against_a_white_background/episode_19/cam_high/frame_000853.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_laptop_box_while_sitting_against_a_white_background/episode_19/cam_high/frame_000025.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_laptop_box_while_sitting_against_a_white_background/episode_19/cam_high/frame_000025.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_laptop_box_while_sitting_against_a_white_background/episode_19/cam_high/frame_000025.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_laptop_box_while_sitting_against_a_white_background/episode_19/cam_high/frame_000050.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_laptop_box_while_sitting_against_a_white_background/episode_19/cam_high/frame_000050.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_laptop_box_while_sitting_against_a_white_background/episode_19/cam_high/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the measuring tape and measure the dimensions of the laptop box while sitting against a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the measuring tape and measure the dimensions of the laptop box while sitting against a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.0, "hop_absolute_value": 3.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Open_the_box_insert_the_watch_then_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting_episode_9-train-bf_000175-af_000200-plus-0000131", "task": "Open the box insert the watch then close the box on a wooden table with a white background while sitting", "image": [ "images/human_egodex/Open_the_box_insert_the_watch_then_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Open_the_box_insert_the_watch_then_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_9/cam_high/frame_000284.jpg", "images/human_egodex/Open_the_box_insert_the_watch_then_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_9/cam_high/frame_000175.jpg", "images/human_egodex/Open_the_box_insert_the_watch_then_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_9/cam_high/frame_000175.jpg", "images/human_egodex/Open_the_box_insert_the_watch_then_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_9/cam_high/frame_000175.jpg", "images/human_egodex/Open_the_box_insert_the_watch_then_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_9/cam_high/frame_000200.jpg", "images/human_egodex/Open_the_box_insert_the_watch_then_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_9/cam_high/frame_000200.jpg", "images/human_egodex/Open_the_box_insert_the_watch_then_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_9/cam_high/frame_000200.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the box insert the watch then close the box on a wooden table with a white background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the box insert the watch then close the box on a wooden table with a white background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Braid_the_three_strands_of_the_ribbon_while_sitting_against_a_brown_background_episode_5-train-bf_000575-af_000675-plus-0000214", "task": "Braid the three strands of the ribbon while sitting against a brown background", "image": [ "images/human_egodex/Braid_the_three_strands_of_the_ribbon_while_sitting_against_a_brown_background/episode_1/cam_high/frame_000000.jpg", "images/human_egodex/Braid_the_three_strands_of_the_ribbon_while_sitting_against_a_brown_background/episode_1/cam_high/frame_000512.jpg", "images/human_egodex/Braid_the_three_strands_of_the_ribbon_while_sitting_against_a_brown_background/episode_5/cam_high/frame_000575.jpg", "images/human_egodex/Braid_the_three_strands_of_the_ribbon_while_sitting_against_a_brown_background/episode_5/cam_high/frame_000575.jpg", "images/human_egodex/Braid_the_three_strands_of_the_ribbon_while_sitting_against_a_brown_background/episode_5/cam_high/frame_000575.jpg", "images/human_egodex/Braid_the_three_strands_of_the_ribbon_while_sitting_against_a_brown_background/episode_5/cam_high/frame_000675.jpg", "images/human_egodex/Braid_the_three_strands_of_the_ribbon_while_sitting_against_a_brown_background/episode_5/cam_high/frame_000675.jpg", "images/human_egodex/Braid_the_three_strands_of_the_ribbon_while_sitting_against_a_brown_background/episode_5/cam_high/frame_000675.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Braid the three strands of the ribbon while sitting against a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Braid the three strands of the ribbon while sitting against a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one_episode_2-train-bf_000225-af_000300-plus-0000133", "task": "Reverse the action by popping out the pressed bubbles on the square fidget Pop It toy one by one", "image": [ "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_2/cam_high/frame_000415.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_2/cam_high/frame_000225.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_2/cam_high/frame_000225.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_2/cam_high/frame_000225.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_2/cam_high/frame_000300.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_2/cam_high/frame_000300.jpg", "images/human_egodex/Reverse_the_action_by_popping_out_the_pressed_bubbles_on_the_square_fidget_Pop_It_toy_one_by_one/episode_2/cam_high/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Reverse the action by popping out the pressed bubbles on the square fidget Pop It toy one by one`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Reverse the action by popping out the pressed bubbles on the square fidget Pop It toy one by one`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 42.9, "hop_absolute_value": 42.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Fold_a_piece_of_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background_episode_1-train-bf_000425-af_000450-plus-0000335", "task": "Fold a piece of paper while sitting at a table with a pink tablecloth and lavender background", "image": [ "images/human_egodex/Fold_a_piece_of_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000000.jpg", "images/human_egodex/Fold_a_piece_of_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000602.jpg", "images/human_egodex/Fold_a_piece_of_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000425.jpg", "images/human_egodex/Fold_a_piece_of_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000425.jpg", "images/human_egodex/Fold_a_piece_of_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000425.jpg", "images/human_egodex/Fold_a_piece_of_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000450.jpg", "images/human_egodex/Fold_a_piece_of_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000450.jpg", "images/human_egodex/Fold_a_piece_of_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold a piece of paper while sitting at a table with a pink tablecloth and lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold a piece of paper while sitting at a table with a pink tablecloth and lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 14.3, "hop_absolute_value": 14.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Place_the_coin_on_top_of_the_thumb_flip_the_coin_catch_it_with_the_other_hand_and_put_it_back_in_position_episode_11-train-bf_000025-af_004225-plus-0000303", "task": "Place the coin on top of the thumb flip the coin catch it with the other hand and put it back in position", "image": [ "images/human_egodex/Place_the_coin_on_top_of_the_thumb_flip_the_coin_catch_it_with_the_other_hand_and_put_it_back_in_position/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Place_the_coin_on_top_of_the_thumb_flip_the_coin_catch_it_with_the_other_hand_and_put_it_back_in_position/episode_11/cam_high/frame_004775.jpg", "images/human_egodex/Place_the_coin_on_top_of_the_thumb_flip_the_coin_catch_it_with_the_other_hand_and_put_it_back_in_position/episode_11/cam_high/frame_000025.jpg", "images/human_egodex/Place_the_coin_on_top_of_the_thumb_flip_the_coin_catch_it_with_the_other_hand_and_put_it_back_in_position/episode_11/cam_high/frame_000025.jpg", "images/human_egodex/Place_the_coin_on_top_of_the_thumb_flip_the_coin_catch_it_with_the_other_hand_and_put_it_back_in_position/episode_11/cam_high/frame_000025.jpg", "images/human_egodex/Place_the_coin_on_top_of_the_thumb_flip_the_coin_catch_it_with_the_other_hand_and_put_it_back_in_position/episode_11/cam_high/frame_004225.jpg", "images/human_egodex/Place_the_coin_on_top_of_the_thumb_flip_the_coin_catch_it_with_the_other_hand_and_put_it_back_in_position/episode_11/cam_high/frame_004225.jpg", "images/human_egodex/Place_the_coin_on_top_of_the_thumb_flip_the_coin_catch_it_with_the_other_hand_and_put_it_back_in_position/episode_11/cam_high/frame_004225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the coin on top of the thumb flip the coin catch it with the other hand and put it back in position`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the coin on top of the thumb flip the coin catch it with the other hand and put it back in position`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 88.4, "hop_absolute_value": 88.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Boil_water_in_pot_take_egg_from_carton_boil_egg_cover_pot_with_lid_then_serve_boiled_egg_in_gray_bowl_placed_on_round_table_episode_10-train-bf_001650-af_001825-plus-0000170", "task": "Boil water in pot take egg from carton boil egg cover pot with lid then serve boiled egg in gray bowl placed on round table", "image": [ "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_cover_pot_with_lid_then_serve_boiled_egg_in_gray_bowl_placed_on_round_table/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_cover_pot_with_lid_then_serve_boiled_egg_in_gray_bowl_placed_on_round_table/episode_7/cam_high/frame_004436.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_cover_pot_with_lid_then_serve_boiled_egg_in_gray_bowl_placed_on_round_table/episode_10/cam_high/frame_001650.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_cover_pot_with_lid_then_serve_boiled_egg_in_gray_bowl_placed_on_round_table/episode_10/cam_high/frame_001650.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_cover_pot_with_lid_then_serve_boiled_egg_in_gray_bowl_placed_on_round_table/episode_10/cam_high/frame_001650.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_cover_pot_with_lid_then_serve_boiled_egg_in_gray_bowl_placed_on_round_table/episode_10/cam_high/frame_001825.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_cover_pot_with_lid_then_serve_boiled_egg_in_gray_bowl_placed_on_round_table/episode_10/cam_high/frame_001825.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_egg_cover_pot_with_lid_then_serve_boiled_egg_in_gray_bowl_placed_on_round_table/episode_10/cam_high/frame_001825.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Boil water in pot take egg from carton boil egg cover pot with lid then serve boiled egg in gray bowl placed on round table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Boil water in pot take egg from carton boil egg cover pot with lid then serve boiled egg in gray bowl placed on round table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.1, "hop_absolute_value": 12.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting_episode_16-train-bf_000400-af_000425-plus-0000376", "task": "Disassemble tiles on a metal table with a lavender background while sitting", "image": [ "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_16/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_16/cam_high/frame_000939.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_16/cam_high/frame_000400.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_16/cam_high/frame_000400.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_16/cam_high/frame_000400.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_16/cam_high/frame_000425.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_16/cam_high/frame_000425.jpg", "images/human_egodex/Disassemble_tiles_on_a_metal_table_with_a_lavender_background_while_sitting/episode_16/cam_high/frame_000425.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble tiles on a metal table with a lavender background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble tiles on a metal table with a lavender background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.8, "hop_absolute_value": 4.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_episode_19-train-bf_000125-af_000200-plus-0000154", "task": "Unscrew the screw from the fixture using an allen key on a pink tablecloth with a lavender background", "image": [ "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background/episode_19/cam_high/frame_000000.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background/episode_19/cam_high/frame_001106.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background/episode_19/cam_high/frame_000125.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background/episode_19/cam_high/frame_000125.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background/episode_19/cam_high/frame_000125.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background/episode_19/cam_high/frame_000200.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background/episode_19/cam_high/frame_000200.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background/episode_19/cam_high/frame_000200.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unscrew the screw from the fixture using an allen key on a pink tablecloth with a lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unscrew the screw from the fixture using an allen key on a pink tablecloth with a lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 7.7, "hop_absolute_value": 7.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unwrapping_a_croissant_from_aluminum_foil_episode_11-train-bf_000325-af_000375-plus-0000459", "task": "Unwrapping a croissant from aluminum foil", "image": [ "images/human_egodex/Unwrapping_a_croissant_from_aluminum_foil/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Unwrapping_a_croissant_from_aluminum_foil/episode_18/cam_high/frame_001805.jpg", "images/human_egodex/Unwrapping_a_croissant_from_aluminum_foil/episode_11/cam_high/frame_000325.jpg", "images/human_egodex/Unwrapping_a_croissant_from_aluminum_foil/episode_11/cam_high/frame_000325.jpg", "images/human_egodex/Unwrapping_a_croissant_from_aluminum_foil/episode_11/cam_high/frame_000325.jpg", "images/human_egodex/Unwrapping_a_croissant_from_aluminum_foil/episode_11/cam_high/frame_000375.jpg", "images/human_egodex/Unwrapping_a_croissant_from_aluminum_foil/episode_11/cam_high/frame_000375.jpg", "images/human_egodex/Unwrapping_a_croissant_from_aluminum_foil/episode_11/cam_high/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unwrapping a croissant from aluminum foil`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unwrapping a croissant from aluminum foil`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 6.7, "hop_absolute_value": 6.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Insert_a_USB_into_the_hairbrush_with_the_left_hand_while_sitting_against_a_blue_background_episode_4-train-bf_000000-af_000125-plus-0000443", "task": "Insert a USB into the hairbrush with the left hand while sitting against a blue background", "image": [ "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_left_hand_while_sitting_against_a_blue_background/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_left_hand_while_sitting_against_a_blue_background/episode_18/cam_high/frame_000877.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_left_hand_while_sitting_against_a_blue_background/episode_4/cam_high/frame_000000.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_left_hand_while_sitting_against_a_blue_background/episode_4/cam_high/frame_000000.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_left_hand_while_sitting_against_a_blue_background/episode_4/cam_high/frame_000000.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_left_hand_while_sitting_against_a_blue_background/episode_4/cam_high/frame_000125.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_left_hand_while_sitting_against_a_blue_background/episode_4/cam_high/frame_000125.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_left_hand_while_sitting_against_a_blue_background/episode_4/cam_high/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert a USB into the hairbrush with the left hand while sitting against a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert a USB into the hairbrush with the left hand while sitting against a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 14.7, "hop_absolute_value": 14.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_stamp_press_it_onto_the_purple_or_pink_ink_pad_stamp_it_onto_the_paper_and_place_it_back_in_its_original_spot_on_the_wooden_tablecloth_episode_0-train-bf_000275-af_000300-plus-0000109", "task": "Pick up the stamp press it onto the purple or pink ink pad stamp it onto the paper and place it back in its original spot on the wooden tablecloth", "image": [ "images/human_egodex/Pick_up_the_stamp_press_it_onto_the_purple_or_pink_ink_pad_stamp_it_onto_the_paper_and_place_it_back_in_its_original_spot_on_the_wooden_tablecloth/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_stamp_press_it_onto_the_purple_or_pink_ink_pad_stamp_it_onto_the_paper_and_place_it_back_in_its_original_spot_on_the_wooden_tablecloth/episode_8/cam_high/frame_000236.jpg", "images/human_egodex/Pick_up_the_stamp_press_it_onto_the_purple_or_pink_ink_pad_stamp_it_onto_the_paper_and_place_it_back_in_its_original_spot_on_the_wooden_tablecloth/episode_0/cam_high/frame_000275.jpg", "images/human_egodex/Pick_up_the_stamp_press_it_onto_the_purple_or_pink_ink_pad_stamp_it_onto_the_paper_and_place_it_back_in_its_original_spot_on_the_wooden_tablecloth/episode_0/cam_high/frame_000275.jpg", "images/human_egodex/Pick_up_the_stamp_press_it_onto_the_purple_or_pink_ink_pad_stamp_it_onto_the_paper_and_place_it_back_in_its_original_spot_on_the_wooden_tablecloth/episode_0/cam_high/frame_000275.jpg", "images/human_egodex/Pick_up_the_stamp_press_it_onto_the_purple_or_pink_ink_pad_stamp_it_onto_the_paper_and_place_it_back_in_its_original_spot_on_the_wooden_tablecloth/episode_0/cam_high/frame_000300.jpg", "images/human_egodex/Pick_up_the_stamp_press_it_onto_the_purple_or_pink_ink_pad_stamp_it_onto_the_paper_and_place_it_back_in_its_original_spot_on_the_wooden_tablecloth/episode_0/cam_high/frame_000300.jpg", "images/human_egodex/Pick_up_the_stamp_press_it_onto_the_purple_or_pink_ink_pad_stamp_it_onto_the_paper_and_place_it_back_in_its_original_spot_on_the_wooden_tablecloth/episode_0/cam_high/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the stamp press it onto the purple or pink ink pad stamp it onto the paper and place it back in its original spot on the wooden tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the stamp press it onto the purple or pink ink pad stamp it onto the paper and place it back in its original spot on the wooden tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background_episode_11-train-bf_000550-af_000425-minus-0000354", "task": "Make a sandwich using breads cheese petty pickles egg lettuce ketchup and mayo on a blue tablecloth with a pink background", "image": [ "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_11/cam_high/frame_000580.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_11/cam_high/frame_000550.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_11/cam_high/frame_000550.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_11/cam_high/frame_000550.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_11/cam_high/frame_000425.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_11/cam_high/frame_000425.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_11/cam_high/frame_000425.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make a sandwich using breads cheese petty pickles egg lettuce ketchup and mayo on a blue tablecloth with a pink background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make a sandwich using breads cheese petty pickles egg lettuce ketchup and mayo on a blue tablecloth with a pink background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -22.7, "hop_absolute_value": 22.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Assemble_the_seahorse_puzzle_on_a_metal_table_while_sitting_episode_11-train-bf_000299-af_000225-minus-0000058", "task": "Assemble the seahorse puzzle on a metal table while sitting", "image": [ "images/human_egodex/Assemble_the_seahorse_puzzle_on_a_metal_table_while_sitting/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Assemble_the_seahorse_puzzle_on_a_metal_table_while_sitting/episode_11/cam_high/frame_000299.jpg", "images/human_egodex/Assemble_the_seahorse_puzzle_on_a_metal_table_while_sitting/episode_11/cam_high/frame_000299.jpg", "images/human_egodex/Assemble_the_seahorse_puzzle_on_a_metal_table_while_sitting/episode_11/cam_high/frame_000299.jpg", "images/human_egodex/Assemble_the_seahorse_puzzle_on_a_metal_table_while_sitting/episode_11/cam_high/frame_000299.jpg", "images/human_egodex/Assemble_the_seahorse_puzzle_on_a_metal_table_while_sitting/episode_11/cam_high/frame_000225.jpg", "images/human_egodex/Assemble_the_seahorse_puzzle_on_a_metal_table_while_sitting/episode_11/cam_high/frame_000225.jpg", "images/human_egodex/Assemble_the_seahorse_puzzle_on_a_metal_table_while_sitting/episode_11/cam_high/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Assemble the seahorse puzzle on a metal table while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Assemble the seahorse puzzle on a metal table while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -18.2, "hop_absolute_value": 18.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_blue_butter_slime_from_the_table_and_place_it_in_the_utensil_drying_rack_episode_3-train-bf_000125-af_000075-minus-0000015", "task": "Pick up the blue butter slime from the table and place it in the utensil drying rack", "image": [ "images/human_egodex/Pick_up_the_blue_butter_slime_from_the_table_and_place_it_in_the_utensil_drying_rack/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_blue_butter_slime_from_the_table_and_place_it_in_the_utensil_drying_rack/episode_3/cam_high/frame_000154.jpg", "images/human_egodex/Pick_up_the_blue_butter_slime_from_the_table_and_place_it_in_the_utensil_drying_rack/episode_3/cam_high/frame_000125.jpg", "images/human_egodex/Pick_up_the_blue_butter_slime_from_the_table_and_place_it_in_the_utensil_drying_rack/episode_3/cam_high/frame_000125.jpg", "images/human_egodex/Pick_up_the_blue_butter_slime_from_the_table_and_place_it_in_the_utensil_drying_rack/episode_3/cam_high/frame_000125.jpg", "images/human_egodex/Pick_up_the_blue_butter_slime_from_the_table_and_place_it_in_the_utensil_drying_rack/episode_3/cam_high/frame_000075.jpg", "images/human_egodex/Pick_up_the_blue_butter_slime_from_the_table_and_place_it_in_the_utensil_drying_rack/episode_3/cam_high/frame_000075.jpg", "images/human_egodex/Pick_up_the_blue_butter_slime_from_the_table_and_place_it_in_the_utensil_drying_rack/episode_3/cam_high/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the blue butter slime from the table and place it in the utensil drying rack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the blue butter slime from the table and place it in the utensil drying rack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Open_the_box_remove_five_duck_objects_and_five_plush_objects_then_close_the_box_on_a_wooden_table_with_a_white_background_episode_9-train-bf_000425-af_000325-minus-0000115", "task": "Open the box remove five duck objects and five plush objects then close the box on a wooden table with a white background", "image": [ "images/human_egodex/Open_the_box_remove_five_duck_objects_and_five_plush_objects_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Open_the_box_remove_five_duck_objects_and_five_plush_objects_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_9/cam_high/frame_000513.jpg", "images/human_egodex/Open_the_box_remove_five_duck_objects_and_five_plush_objects_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_9/cam_high/frame_000425.jpg", "images/human_egodex/Open_the_box_remove_five_duck_objects_and_five_plush_objects_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_9/cam_high/frame_000425.jpg", "images/human_egodex/Open_the_box_remove_five_duck_objects_and_five_plush_objects_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_9/cam_high/frame_000425.jpg", "images/human_egodex/Open_the_box_remove_five_duck_objects_and_five_plush_objects_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_9/cam_high/frame_000325.jpg", "images/human_egodex/Open_the_box_remove_five_duck_objects_and_five_plush_objects_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_9/cam_high/frame_000325.jpg", "images/human_egodex/Open_the_box_remove_five_duck_objects_and_five_plush_objects_then_close_the_box_on_a_wooden_table_with_a_white_background/episode_9/cam_high/frame_000325.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the box remove five duck objects and five plush objects then close the box on a wooden table with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the box remove five duck objects and five plush objects then close the box on a wooden table with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -23.5, "hop_absolute_value": 23.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Insert_a_tennis_ball_into_a_group_of_three_tennis_balls_on_a_pink_tablecloth_episode_3-train-bf_000075-af_000000-minus-0000002", "task": "Insert a tennis ball into a group of three tennis balls on a pink tablecloth", "image": [ "images/human_egodex/Insert_a_tennis_ball_into_a_group_of_three_tennis_balls_on_a_pink_tablecloth/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Insert_a_tennis_ball_into_a_group_of_three_tennis_balls_on_a_pink_tablecloth/episode_3/cam_high/frame_000433.jpg", "images/human_egodex/Insert_a_tennis_ball_into_a_group_of_three_tennis_balls_on_a_pink_tablecloth/episode_3/cam_high/frame_000075.jpg", "images/human_egodex/Insert_a_tennis_ball_into_a_group_of_three_tennis_balls_on_a_pink_tablecloth/episode_3/cam_high/frame_000075.jpg", "images/human_egodex/Insert_a_tennis_ball_into_a_group_of_three_tennis_balls_on_a_pink_tablecloth/episode_3/cam_high/frame_000075.jpg", "images/human_egodex/Insert_a_tennis_ball_into_a_group_of_three_tennis_balls_on_a_pink_tablecloth/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Insert_a_tennis_ball_into_a_group_of_three_tennis_balls_on_a_pink_tablecloth/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Insert_a_tennis_ball_into_a_group_of_three_tennis_balls_on_a_pink_tablecloth/episode_3/cam_high/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert a tennis ball into a group of three tennis balls on a pink tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert a tennis ball into a group of three tennis balls on a pink tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unwrap_the_sandwich_from_the_aluminum_foil_and_place_it_back_on_the_gray_plate_episode_1-train-bf_000650-af_000500-minus-0000413", "task": "Unwrap the sandwich from the aluminum foil and place it back on the gray plate", "image": [ "images/human_egodex/Unwrap_the_sandwich_from_the_aluminum_foil_and_place_it_back_on_the_gray_plate/episode_19/cam_high/frame_000000.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_aluminum_foil_and_place_it_back_on_the_gray_plate/episode_19/cam_high/frame_000843.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_aluminum_foil_and_place_it_back_on_the_gray_plate/episode_1/cam_high/frame_000650.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_aluminum_foil_and_place_it_back_on_the_gray_plate/episode_1/cam_high/frame_000650.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_aluminum_foil_and_place_it_back_on_the_gray_plate/episode_1/cam_high/frame_000650.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_aluminum_foil_and_place_it_back_on_the_gray_plate/episode_1/cam_high/frame_000500.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_aluminum_foil_and_place_it_back_on_the_gray_plate/episode_1/cam_high/frame_000500.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_aluminum_foil_and_place_it_back_on_the_gray_plate/episode_1/cam_high/frame_000500.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unwrap the sandwich from the aluminum foil and place it back on the gray plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unwrap the sandwich from the aluminum foil and place it back on the gray plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -23.1, "hop_absolute_value": 23.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Wrap_a_croissant_with_plastic_wrap_episode_2-train-bf_000375-af_000250-minus-0000346", "task": "Wrap a croissant with plastic wrap", "image": [ "images/human_egodex/Wrap_a_croissant_with_plastic_wrap/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Wrap_a_croissant_with_plastic_wrap/episode_2/cam_high/frame_000533.jpg", "images/human_egodex/Wrap_a_croissant_with_plastic_wrap/episode_2/cam_high/frame_000375.jpg", "images/human_egodex/Wrap_a_croissant_with_plastic_wrap/episode_2/cam_high/frame_000375.jpg", "images/human_egodex/Wrap_a_croissant_with_plastic_wrap/episode_2/cam_high/frame_000375.jpg", "images/human_egodex/Wrap_a_croissant_with_plastic_wrap/episode_2/cam_high/frame_000250.jpg", "images/human_egodex/Wrap_a_croissant_with_plastic_wrap/episode_2/cam_high/frame_000250.jpg", "images/human_egodex/Wrap_a_croissant_with_plastic_wrap/episode_2/cam_high/frame_000250.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Wrap a croissant with plastic wrap`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Wrap a croissant with plastic wrap`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_blue_and_red_tableware_including_plates_forks_spoons_knives_and_cups_from_the_green_tablecloth_and_fold_it_episode_17-train-bf_001250-af_001100-minus-0000369", "task": "Remove the blue and red tableware including plates forks spoons knives and cups from the green tablecloth and fold it", "image": [ "images/human_egodex/Remove_the_blue_and_red_tableware_including_plates_forks_spoons_knives_and_cups_from_the_green_tablecloth_and_fold_it/episode_17/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_blue_and_red_tableware_including_plates_forks_spoons_knives_and_cups_from_the_green_tablecloth_and_fold_it/episode_17/cam_high/frame_001286.jpg", "images/human_egodex/Remove_the_blue_and_red_tableware_including_plates_forks_spoons_knives_and_cups_from_the_green_tablecloth_and_fold_it/episode_17/cam_high/frame_001250.jpg", "images/human_egodex/Remove_the_blue_and_red_tableware_including_plates_forks_spoons_knives_and_cups_from_the_green_tablecloth_and_fold_it/episode_17/cam_high/frame_001250.jpg", "images/human_egodex/Remove_the_blue_and_red_tableware_including_plates_forks_spoons_knives_and_cups_from_the_green_tablecloth_and_fold_it/episode_17/cam_high/frame_001250.jpg", "images/human_egodex/Remove_the_blue_and_red_tableware_including_plates_forks_spoons_knives_and_cups_from_the_green_tablecloth_and_fold_it/episode_17/cam_high/frame_001100.jpg", "images/human_egodex/Remove_the_blue_and_red_tableware_including_plates_forks_spoons_knives_and_cups_from_the_green_tablecloth_and_fold_it/episode_17/cam_high/frame_001100.jpg", "images/human_egodex/Remove_the_blue_and_red_tableware_including_plates_forks_spoons_knives_and_cups_from_the_green_tablecloth_and_fold_it/episode_17/cam_high/frame_001100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the blue and red tableware including plates forks spoons knives and cups from the green tablecloth and fold it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the blue and red tableware including plates forks spoons knives and cups from the green tablecloth and fold it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.0, "hop_absolute_value": 12.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Take_toothbrush_from_tall_coffee_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_tall_coffee_cup_episode_7-train-bf_000925-af_000900-minus-0000321", "task": "Take toothbrush from tall coffee cup unscrew toothpaste cap put toothpaste on toothbrush screw cap back on toothpaste put toothpaste down and put toothbrush back into tall coffee cup", "image": [ "images/human_egodex/Take_toothbrush_from_tall_coffee_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_tall_coffee_cup/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Take_toothbrush_from_tall_coffee_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_tall_coffee_cup/episode_7/cam_high/frame_000987.jpg", "images/human_egodex/Take_toothbrush_from_tall_coffee_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_tall_coffee_cup/episode_7/cam_high/frame_000925.jpg", "images/human_egodex/Take_toothbrush_from_tall_coffee_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_tall_coffee_cup/episode_7/cam_high/frame_000925.jpg", "images/human_egodex/Take_toothbrush_from_tall_coffee_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_tall_coffee_cup/episode_7/cam_high/frame_000925.jpg", "images/human_egodex/Take_toothbrush_from_tall_coffee_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_tall_coffee_cup/episode_7/cam_high/frame_000900.jpg", "images/human_egodex/Take_toothbrush_from_tall_coffee_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_tall_coffee_cup/episode_7/cam_high/frame_000900.jpg", "images/human_egodex/Take_toothbrush_from_tall_coffee_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_tall_coffee_cup/episode_7/cam_high/frame_000900.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take toothbrush from tall coffee cup unscrew toothpaste cap put toothpaste on toothbrush screw cap back on toothpaste put toothpaste down and put toothbrush back into tall coffee cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take toothbrush from tall coffee cup unscrew toothpaste cap put toothpaste on toothbrush screw cap back on toothpaste put toothpaste down and put toothbrush back into tall coffee cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.7, "hop_absolute_value": 2.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_brown_egg_from_the_clear_tupperware_with_the_right_hand_and_place_it_into_the_orange_bowl_episode_15-train-bf_000112-af_000075-minus-0000009", "task": "Pick up the brown egg from the clear tupperware with the right hand and place it into the orange bowl", "image": [ "images/human_egodex/Pick_up_the_brown_egg_from_the_clear_tupperware_with_the_right_hand_and_place_it_into_the_orange_bowl/episode_15/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_brown_egg_from_the_clear_tupperware_with_the_right_hand_and_place_it_into_the_orange_bowl/episode_15/cam_high/frame_000112.jpg", "images/human_egodex/Pick_up_the_brown_egg_from_the_clear_tupperware_with_the_right_hand_and_place_it_into_the_orange_bowl/episode_15/cam_high/frame_000112.jpg", "images/human_egodex/Pick_up_the_brown_egg_from_the_clear_tupperware_with_the_right_hand_and_place_it_into_the_orange_bowl/episode_15/cam_high/frame_000112.jpg", "images/human_egodex/Pick_up_the_brown_egg_from_the_clear_tupperware_with_the_right_hand_and_place_it_into_the_orange_bowl/episode_15/cam_high/frame_000112.jpg", "images/human_egodex/Pick_up_the_brown_egg_from_the_clear_tupperware_with_the_right_hand_and_place_it_into_the_orange_bowl/episode_15/cam_high/frame_000075.jpg", "images/human_egodex/Pick_up_the_brown_egg_from_the_clear_tupperware_with_the_right_hand_and_place_it_into_the_orange_bowl/episode_15/cam_high/frame_000075.jpg", "images/human_egodex/Pick_up_the_brown_egg_from_the_clear_tupperware_with_the_right_hand_and_place_it_into_the_orange_bowl/episode_15/cam_high/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the brown egg from the clear tupperware with the right hand and place it into the orange bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the brown egg from the clear tupperware with the right hand and place it into the orange bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table_episode_3-train-bf_000700-af_000050-minus-0000006", "task": "Remove the pokemon card from the black card sleeve and place it back onto the metal table", "image": [ "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_9/cam_high/frame_002321.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_3/cam_high/frame_000700.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_3/cam_high/frame_000700.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_3/cam_high/frame_000700.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_3/cam_high/frame_000050.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_3/cam_high/frame_000050.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_3/cam_high/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the pokemon card from the black card sleeve and place it back onto the metal table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the pokemon card from the black card sleeve and place it back onto the metal table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -92.9, "hop_absolute_value": 92.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Collect_rubber_ducks_and_towels_scattered_across_the_metal_tablecloth_and_then_throw_them_into_a_box_episode_1-train-bf_000575-af_000350-minus-0000112", "task": "Collect rubber ducks and towels scattered across the metal tablecloth and then throw them into a box", "image": [ "images/human_egodex/Collect_rubber_ducks_and_towels_scattered_across_the_metal_tablecloth_and_then_throw_them_into_a_box/episode_16/cam_high/frame_000000.jpg", "images/human_egodex/Collect_rubber_ducks_and_towels_scattered_across_the_metal_tablecloth_and_then_throw_them_into_a_box/episode_16/cam_high/frame_000475.jpg", "images/human_egodex/Collect_rubber_ducks_and_towels_scattered_across_the_metal_tablecloth_and_then_throw_them_into_a_box/episode_1/cam_high/frame_000575.jpg", "images/human_egodex/Collect_rubber_ducks_and_towels_scattered_across_the_metal_tablecloth_and_then_throw_them_into_a_box/episode_1/cam_high/frame_000575.jpg", "images/human_egodex/Collect_rubber_ducks_and_towels_scattered_across_the_metal_tablecloth_and_then_throw_them_into_a_box/episode_1/cam_high/frame_000575.jpg", "images/human_egodex/Collect_rubber_ducks_and_towels_scattered_across_the_metal_tablecloth_and_then_throw_them_into_a_box/episode_1/cam_high/frame_000350.jpg", "images/human_egodex/Collect_rubber_ducks_and_towels_scattered_across_the_metal_tablecloth_and_then_throw_them_into_a_box/episode_1/cam_high/frame_000350.jpg", "images/human_egodex/Collect_rubber_ducks_and_towels_scattered_across_the_metal_tablecloth_and_then_throw_them_into_a_box/episode_1/cam_high/frame_000350.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Collect rubber ducks and towels scattered across the metal tablecloth and then throw them into a box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Collect rubber ducks and towels scattered across the metal tablecloth and then throw them into a box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -39.1, "hop_absolute_value": 39.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_two_clear_tupperware_from_a_wooden_table_with_the_right_hand_and_place_them_on_a_shelf_episode_10-train-bf_000425-af_000025-minus-0000014", "task": "Pick up two clear tupperware from a wooden table with the right hand and place them on a shelf", "image": [ "images/human_egodex/Pick_up_two_clear_tupperware_from_a_wooden_table_with_the_right_hand_and_place_them_on_a_shelf/episode_10/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_two_clear_tupperware_from_a_wooden_table_with_the_right_hand_and_place_them_on_a_shelf/episode_10/cam_high/frame_000523.jpg", "images/human_egodex/Pick_up_two_clear_tupperware_from_a_wooden_table_with_the_right_hand_and_place_them_on_a_shelf/episode_10/cam_high/frame_000425.jpg", "images/human_egodex/Pick_up_two_clear_tupperware_from_a_wooden_table_with_the_right_hand_and_place_them_on_a_shelf/episode_10/cam_high/frame_000425.jpg", "images/human_egodex/Pick_up_two_clear_tupperware_from_a_wooden_table_with_the_right_hand_and_place_them_on_a_shelf/episode_10/cam_high/frame_000425.jpg", "images/human_egodex/Pick_up_two_clear_tupperware_from_a_wooden_table_with_the_right_hand_and_place_them_on_a_shelf/episode_10/cam_high/frame_000025.jpg", "images/human_egodex/Pick_up_two_clear_tupperware_from_a_wooden_table_with_the_right_hand_and_place_them_on_a_shelf/episode_10/cam_high/frame_000025.jpg", "images/human_egodex/Pick_up_two_clear_tupperware_from_a_wooden_table_with_the_right_hand_and_place_them_on_a_shelf/episode_10/cam_high/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up two clear tupperware from a wooden table with the right hand and place them on a shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up two clear tupperware from a wooden table with the right hand and place them on a shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -94.1, "hop_absolute_value": 94.1, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve_episode_2-train-bf_001650-af_000650-minus-0000083", "task": "Pick up a Pokemon or Vanguard card from the metal table and insert it into a clear binder sleeve", "image": [ "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_14/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_14/cam_high/frame_001898.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_2/cam_high/frame_001650.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_2/cam_high/frame_001650.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_2/cam_high/frame_001650.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_2/cam_high/frame_000650.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_2/cam_high/frame_000650.jpg", "images/human_egodex/Pick_up_a_Pokemon_or_Vanguard_card_from_the_metal_table_and_insert_it_into_a_clear_binder_sleeve/episode_2/cam_high/frame_000650.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up a Pokemon or Vanguard card from the metal table and insert it into a clear binder sleeve`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up a Pokemon or Vanguard card from the metal table and insert it into a clear binder sleeve`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -60.6, "hop_absolute_value": 60.6, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate_episode_8-train-bf_000700-af_000625-minus-0000368", "task": "Unwrap the sandwich from the plastic wrap and place it back on the gray plate", "image": [ "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000997.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000700.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000700.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000700.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000625.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000625.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_8/cam_high/frame_000625.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unwrap the sandwich from the plastic wrap and place it back on the gray plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unwrap the sandwich from the plastic wrap and place it back on the gray plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.7, "hop_absolute_value": 10.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Open_the_box_remove_the_pad_and_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting_episode_1-train-bf_000138-af_000025-minus-0000005", "task": "Open the box remove the pad and close the box on a wooden table with a white background while sitting", "image": [ "images/human_egodex/Open_the_box_remove_the_pad_and_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_1/cam_high/frame_000000.jpg", "images/human_egodex/Open_the_box_remove_the_pad_and_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_1/cam_high/frame_000138.jpg", "images/human_egodex/Open_the_box_remove_the_pad_and_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_1/cam_high/frame_000138.jpg", "images/human_egodex/Open_the_box_remove_the_pad_and_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_1/cam_high/frame_000138.jpg", "images/human_egodex/Open_the_box_remove_the_pad_and_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_1/cam_high/frame_000138.jpg", "images/human_egodex/Open_the_box_remove_the_pad_and_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_1/cam_high/frame_000025.jpg", "images/human_egodex/Open_the_box_remove_the_pad_and_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_1/cam_high/frame_000025.jpg", "images/human_egodex/Open_the_box_remove_the_pad_and_close_the_box_on_a_wooden_table_with_a_white_background_while_sitting/episode_1/cam_high/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the box remove the pad and close the box on a wooden table with a white background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the box remove the pad and close the box on a wooden table with a white background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Open_the_fridge_door_remove_six_food_items_from_the_fridge_place_the_food_items_onto_the_table_and_close_the_fridge_door_episode_3-train-bf_000575-af_000350-minus-0000115", "task": "Open the fridge door remove six food items from the fridge place the food items onto the table and close the fridge door", "image": [ "images/human_egodex/Open_the_fridge_door_remove_six_food_items_from_the_fridge_place_the_food_items_onto_the_table_and_close_the_fridge_door/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Open_the_fridge_door_remove_six_food_items_from_the_fridge_place_the_food_items_onto_the_table_and_close_the_fridge_door/episode_3/cam_high/frame_000689.jpg", "images/human_egodex/Open_the_fridge_door_remove_six_food_items_from_the_fridge_place_the_food_items_onto_the_table_and_close_the_fridge_door/episode_3/cam_high/frame_000575.jpg", "images/human_egodex/Open_the_fridge_door_remove_six_food_items_from_the_fridge_place_the_food_items_onto_the_table_and_close_the_fridge_door/episode_3/cam_high/frame_000575.jpg", "images/human_egodex/Open_the_fridge_door_remove_six_food_items_from_the_fridge_place_the_food_items_onto_the_table_and_close_the_fridge_door/episode_3/cam_high/frame_000575.jpg", "images/human_egodex/Open_the_fridge_door_remove_six_food_items_from_the_fridge_place_the_food_items_onto_the_table_and_close_the_fridge_door/episode_3/cam_high/frame_000350.jpg", "images/human_egodex/Open_the_fridge_door_remove_six_food_items_from_the_fridge_place_the_food_items_onto_the_table_and_close_the_fridge_door/episode_3/cam_high/frame_000350.jpg", "images/human_egodex/Open_the_fridge_door_remove_six_food_items_from_the_fridge_place_the_food_items_onto_the_table_and_close_the_fridge_door/episode_3/cam_high/frame_000350.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the fridge door remove six food items from the fridge place the food items onto the table and close the fridge door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the fridge door remove six food items from the fridge place the food items onto the table and close the fridge door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -39.1, "hop_absolute_value": 39.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_egg_from_bowl_place_back_into_egg_carton_and_remove_pot_lid_episode_3-train-bf_000492-af_000375-minus-0000289", "task": "Remove egg from bowl place back into egg carton and remove pot lid", "image": [ "images/human_egodex/Remove_egg_from_bowl_place_back_into_egg_carton_and_remove_pot_lid/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Remove_egg_from_bowl_place_back_into_egg_carton_and_remove_pot_lid/episode_7/cam_high/frame_000668.jpg", "images/human_egodex/Remove_egg_from_bowl_place_back_into_egg_carton_and_remove_pot_lid/episode_3/cam_high/frame_000492.jpg", "images/human_egodex/Remove_egg_from_bowl_place_back_into_egg_carton_and_remove_pot_lid/episode_3/cam_high/frame_000492.jpg", "images/human_egodex/Remove_egg_from_bowl_place_back_into_egg_carton_and_remove_pot_lid/episode_3/cam_high/frame_000492.jpg", "images/human_egodex/Remove_egg_from_bowl_place_back_into_egg_carton_and_remove_pot_lid/episode_3/cam_high/frame_000375.jpg", "images/human_egodex/Remove_egg_from_bowl_place_back_into_egg_carton_and_remove_pot_lid/episode_3/cam_high/frame_000375.jpg", "images/human_egodex/Remove_egg_from_bowl_place_back_into_egg_carton_and_remove_pot_lid/episode_3/cam_high/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove egg from bowl place back into egg carton and remove pot lid`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove egg from bowl place back into egg carton and remove pot lid`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -21.1, "hop_absolute_value": 21.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background_episode_18-train-bf_000900-af_000850-minus-0000363", "task": "Make a sandwich using breads lettuce cucumbers onion tomato beef and mustard while sitting at a table with a blue background", "image": [ "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background/episode_18/cam_high/frame_000935.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background/episode_18/cam_high/frame_000900.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background/episode_18/cam_high/frame_000900.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background/episode_18/cam_high/frame_000900.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background/episode_18/cam_high/frame_000850.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background/episode_18/cam_high/frame_000850.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background/episode_18/cam_high/frame_000850.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make a sandwich using breads lettuce cucumbers onion tomato beef and mustard while sitting at a table with a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make a sandwich using breads lettuce cucumbers onion tomato beef and mustard while sitting at a table with a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.6, "hop_absolute_value": 5.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Assemble_the_leg_onto_the_furniture_bench_stool_on_a_wooden_table_with_a_red_background_while_sitting_episode_6-train-bf_000278-af_000150-minus-0000039", "task": "Assemble the leg onto the furniture bench stool on a wooden table with a red background while sitting", "image": [ "images/human_egodex/Assemble_the_leg_onto_the_furniture_bench_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Assemble_the_leg_onto_the_furniture_bench_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_6/cam_high/frame_000278.jpg", "images/human_egodex/Assemble_the_leg_onto_the_furniture_bench_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_6/cam_high/frame_000278.jpg", "images/human_egodex/Assemble_the_leg_onto_the_furniture_bench_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_6/cam_high/frame_000278.jpg", "images/human_egodex/Assemble_the_leg_onto_the_furniture_bench_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_6/cam_high/frame_000278.jpg", "images/human_egodex/Assemble_the_leg_onto_the_furniture_bench_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_6/cam_high/frame_000150.jpg", "images/human_egodex/Assemble_the_leg_onto_the_furniture_bench_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_6/cam_high/frame_000150.jpg", "images/human_egodex/Assemble_the_leg_onto_the_furniture_bench_stool_on_a_wooden_table_with_a_red_background_while_sitting/episode_6/cam_high/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Assemble the leg onto the furniture bench stool on a wooden table with a red background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Assemble the leg onto the furniture bench stool on a wooden table with a red background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -45.5, "hop_absolute_value": 45.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth_episode_17-train-bf_000700-af_000075-minus-0000036", "task": "Make a sandwich using breads lettuce cucumbers onion tomato beef and mustard while sitting at a table with a blue background and no tablecloth", "image": [ "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_4/cam_high/frame_000000.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_4/cam_high/frame_000613.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_17/cam_high/frame_000700.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_17/cam_high/frame_000700.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_17/cam_high/frame_000700.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_17/cam_high/frame_000075.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_17/cam_high/frame_000075.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_onion_tomato_beef_and_mustard_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_17/cam_high/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make a sandwich using breads lettuce cucumbers onion tomato beef and mustard while sitting at a table with a blue background and no tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make a sandwich using breads lettuce cucumbers onion tomato beef and mustard while sitting at a table with a blue background and no tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -89.3, "hop_absolute_value": 89.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unscrew_the_cap_from_the_Gatorade_bottle_placed_on_the_green_tablecloth_episode_0-train-bf_000092-af_000000-minus-0000002", "task": "Unscrew the cap from the Gatorade bottle placed on the green tablecloth", "image": [ "images/human_egodex/Unscrew_the_cap_from_the_Gatorade_bottle_placed_on_the_green_tablecloth/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Unscrew_the_cap_from_the_Gatorade_bottle_placed_on_the_green_tablecloth/episode_0/cam_high/frame_000092.jpg", "images/human_egodex/Unscrew_the_cap_from_the_Gatorade_bottle_placed_on_the_green_tablecloth/episode_0/cam_high/frame_000092.jpg", "images/human_egodex/Unscrew_the_cap_from_the_Gatorade_bottle_placed_on_the_green_tablecloth/episode_0/cam_high/frame_000092.jpg", "images/human_egodex/Unscrew_the_cap_from_the_Gatorade_bottle_placed_on_the_green_tablecloth/episode_0/cam_high/frame_000092.jpg", "images/human_egodex/Unscrew_the_cap_from_the_Gatorade_bottle_placed_on_the_green_tablecloth/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Unscrew_the_cap_from_the_Gatorade_bottle_placed_on_the_green_tablecloth/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Unscrew_the_cap_from_the_Gatorade_bottle_placed_on_the_green_tablecloth/episode_0/cam_high/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unscrew the cap from the Gatorade bottle placed on the green tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unscrew the cap from the Gatorade bottle placed on the green tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting_episode_1-train-bf_000675-af_000600-minus-0000320", "task": "Unscrew the screw from the fixture using an allen key on a pink tablecloth with a lavender background while sitting", "image": [ "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_15/cam_high/frame_000000.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_15/cam_high/frame_000786.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_1/cam_high/frame_000675.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_1/cam_high/frame_000675.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_1/cam_high/frame_000675.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_1/cam_high/frame_000600.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_1/cam_high/frame_000600.jpg", "images/human_egodex/Unscrew_the_screw_from_the_fixture_using_an_allen_key_on_a_pink_tablecloth_with_a_lavender_background_while_sitting/episode_1/cam_high/frame_000600.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unscrew the screw from the fixture using an allen key on a pink tablecloth with a lavender background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unscrew the screw from the fixture using an allen key on a pink tablecloth with a lavender background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Open_the_case_remove_the_pad_and_plush_from_the_case_then_close_the_case_on_a_wooden_table_with_a_brown_background_episode_2-train-bf_000075-af_000025-minus-0000007", "task": "Open the case remove the pad and plush from the case then close the case on a wooden table with a brown background", "image": [ "images/human_egodex/Open_the_case_remove_the_pad_and_plush_from_the_case_then_close_the_case_on_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Open_the_case_remove_the_pad_and_plush_from_the_case_then_close_the_case_on_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000140.jpg", "images/human_egodex/Open_the_case_remove_the_pad_and_plush_from_the_case_then_close_the_case_on_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000075.jpg", "images/human_egodex/Open_the_case_remove_the_pad_and_plush_from_the_case_then_close_the_case_on_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000075.jpg", "images/human_egodex/Open_the_case_remove_the_pad_and_plush_from_the_case_then_close_the_case_on_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000075.jpg", "images/human_egodex/Open_the_case_remove_the_pad_and_plush_from_the_case_then_close_the_case_on_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000025.jpg", "images/human_egodex/Open_the_case_remove_the_pad_and_plush_from_the_case_then_close_the_case_on_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000025.jpg", "images/human_egodex/Open_the_case_remove_the_pad_and_plush_from_the_case_then_close_the_case_on_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the case remove the pad and plush from the case then close the case on a wooden table with a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the case remove the pad and plush from the case then close the case on a wooden table with a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unbraid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background_episode_13-train-bf_000625-af_000200-minus-0000083", "task": "Unbraid the strands of the fishtail yarn while sitting against a brown background", "image": [ "images/human_egodex/Unbraid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_13/cam_high/frame_000000.jpg", "images/human_egodex/Unbraid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_13/cam_high/frame_000787.jpg", "images/human_egodex/Unbraid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_13/cam_high/frame_000625.jpg", "images/human_egodex/Unbraid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_13/cam_high/frame_000625.jpg", "images/human_egodex/Unbraid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_13/cam_high/frame_000625.jpg", "images/human_egodex/Unbraid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_13/cam_high/frame_000200.jpg", "images/human_egodex/Unbraid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_13/cam_high/frame_000200.jpg", "images/human_egodex/Unbraid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_13/cam_high/frame_000200.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unbraid the strands of the fishtail yarn while sitting against a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unbraid the strands of the fishtail yarn while sitting against a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -68.0, "hop_absolute_value": 68.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Plug_the_white_phone_into_the_lightning_cable_on_the_blue_tablecloth_episode_5-train-bf_000100-af_000000-minus-0000003", "task": "Plug the white phone into the lightning cable on the blue tablecloth", "image": [ "images/human_egodex/Plug_the_white_phone_into_the_lightning_cable_on_the_blue_tablecloth/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Plug_the_white_phone_into_the_lightning_cable_on_the_blue_tablecloth/episode_5/cam_high/frame_000151.jpg", "images/human_egodex/Plug_the_white_phone_into_the_lightning_cable_on_the_blue_tablecloth/episode_5/cam_high/frame_000100.jpg", "images/human_egodex/Plug_the_white_phone_into_the_lightning_cable_on_the_blue_tablecloth/episode_5/cam_high/frame_000100.jpg", "images/human_egodex/Plug_the_white_phone_into_the_lightning_cable_on_the_blue_tablecloth/episode_5/cam_high/frame_000100.jpg", "images/human_egodex/Plug_the_white_phone_into_the_lightning_cable_on_the_blue_tablecloth/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Plug_the_white_phone_into_the_lightning_cable_on_the_blue_tablecloth/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Plug_the_white_phone_into_the_lightning_cable_on_the_blue_tablecloth/episode_5/cam_high/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Plug the white phone into the lightning cable on the blue tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Plug the white phone into the lightning cable on the blue tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_a_card_from_the_metal_table_and_insert_it_into_a_black_card_sleeve_episode_2-train-bf_001450-af_001225-minus-0000143", "task": "Pick up a card from the metal table and insert it into a black card sleeve", "image": [ "images/human_egodex/Pick_up_a_card_from_the_metal_table_and_insert_it_into_a_black_card_sleeve/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_a_card_from_the_metal_table_and_insert_it_into_a_black_card_sleeve/episode_5/cam_high/frame_001180.jpg", "images/human_egodex/Pick_up_a_card_from_the_metal_table_and_insert_it_into_a_black_card_sleeve/episode_2/cam_high/frame_001450.jpg", "images/human_egodex/Pick_up_a_card_from_the_metal_table_and_insert_it_into_a_black_card_sleeve/episode_2/cam_high/frame_001450.jpg", "images/human_egodex/Pick_up_a_card_from_the_metal_table_and_insert_it_into_a_black_card_sleeve/episode_2/cam_high/frame_001450.jpg", "images/human_egodex/Pick_up_a_card_from_the_metal_table_and_insert_it_into_a_black_card_sleeve/episode_2/cam_high/frame_001225.jpg", "images/human_egodex/Pick_up_a_card_from_the_metal_table_and_insert_it_into_a_black_card_sleeve/episode_2/cam_high/frame_001225.jpg", "images/human_egodex/Pick_up_a_card_from_the_metal_table_and_insert_it_into_a_black_card_sleeve/episode_2/cam_high/frame_001225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up a card from the metal table and insert it into a black card sleeve`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up a card from the metal table and insert it into a black card sleeve`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -15.5, "hop_absolute_value": 15.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Fold_a_small_tshirt_on_a_green_tablecloth_episode_2-train-bf_000375-af_000000-minus-0000013", "task": "Fold a small tshirt on a green tablecloth", "image": [ "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth/episode_2/cam_high/frame_000585.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth/episode_2/cam_high/frame_000375.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth/episode_2/cam_high/frame_000375.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth/episode_2/cam_high/frame_000375.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth/episode_2/cam_high/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold a small tshirt on a green tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold a small tshirt on a green tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Fold_the_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background_episode_1-train-bf_000200-af_000100-minus-0000040", "task": "Fold the paper while sitting at a table with a pink tablecloth and lavender background", "image": [ "images/human_egodex/Fold_the_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Fold_the_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000333.jpg", "images/human_egodex/Fold_the_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000200.jpg", "images/human_egodex/Fold_the_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000200.jpg", "images/human_egodex/Fold_the_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000200.jpg", "images/human_egodex/Fold_the_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000100.jpg", "images/human_egodex/Fold_the_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000100.jpg", "images/human_egodex/Fold_the_paper_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_1/cam_high/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold the paper while sitting at a table with a pink tablecloth and lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold the paper while sitting at a table with a pink tablecloth and lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Place_lids_onto_four_cups_on_a_wooden_table_with_a_red_background_while_sitting_episode_11-train-bf_000400-af_000175-minus-0000075", "task": "Place lids onto four cups on a wooden table with a red background while sitting", "image": [ "images/human_egodex/Place_lids_onto_four_cups_on_a_wooden_table_with_a_red_background_while_sitting/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Place_lids_onto_four_cups_on_a_wooden_table_with_a_red_background_while_sitting/episode_11/cam_high/frame_000489.jpg", "images/human_egodex/Place_lids_onto_four_cups_on_a_wooden_table_with_a_red_background_while_sitting/episode_11/cam_high/frame_000400.jpg", "images/human_egodex/Place_lids_onto_four_cups_on_a_wooden_table_with_a_red_background_while_sitting/episode_11/cam_high/frame_000400.jpg", "images/human_egodex/Place_lids_onto_four_cups_on_a_wooden_table_with_a_red_background_while_sitting/episode_11/cam_high/frame_000400.jpg", "images/human_egodex/Place_lids_onto_four_cups_on_a_wooden_table_with_a_red_background_while_sitting/episode_11/cam_high/frame_000175.jpg", "images/human_egodex/Place_lids_onto_four_cups_on_a_wooden_table_with_a_red_background_while_sitting/episode_11/cam_high/frame_000175.jpg", "images/human_egodex/Place_lids_onto_four_cups_on_a_wooden_table_with_a_red_background_while_sitting/episode_11/cam_high/frame_000175.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place lids onto four cups on a wooden table with a red background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place lids onto four cups on a wooden table with a red background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -56.2, "hop_absolute_value": 56.2, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_puzzle3_pieces_from_the_wooden_table_with_the_right_hand_and_assemble_the_puzzle_episode_5-train-bf_000275-af_000200-minus-0000091", "task": "Pick up the puzzle3 pieces from the wooden table with the right hand and assemble the puzzle", "image": [ "images/human_egodex/Pick_up_the_puzzle3_pieces_from_the_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_puzzle3_pieces_from_the_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000391.jpg", "images/human_egodex/Pick_up_the_puzzle3_pieces_from_the_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000275.jpg", "images/human_egodex/Pick_up_the_puzzle3_pieces_from_the_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000275.jpg", "images/human_egodex/Pick_up_the_puzzle3_pieces_from_the_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000275.jpg", "images/human_egodex/Pick_up_the_puzzle3_pieces_from_the_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000200.jpg", "images/human_egodex/Pick_up_the_puzzle3_pieces_from_the_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000200.jpg", "images/human_egodex/Pick_up_the_puzzle3_pieces_from_the_wooden_table_with_the_right_hand_and_assemble_the_puzzle/episode_5/cam_high/frame_000200.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the puzzle3 pieces from the wooden table with the right hand and assemble the puzzle`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the puzzle3 pieces from the wooden table with the right hand and assemble the puzzle`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -27.3, "hop_absolute_value": 27.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table_episode_13-train-bf_001125-af_001100-minus-0000431", "task": "Remove the pokemon card from the black card sleeve and place it back onto the metal table", "image": [ "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_13/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_13/cam_high/frame_002501.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_13/cam_high/frame_001125.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_13/cam_high/frame_001125.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_13/cam_high/frame_001125.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_13/cam_high/frame_001100.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_13/cam_high/frame_001100.jpg", "images/human_egodex/Remove_the_pokemon_card_from_the_black_card_sleeve_and_place_it_back_onto_the_metal_table/episode_13/cam_high/frame_001100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the pokemon card from the black card sleeve and place it back onto the metal table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the pokemon card from the black card sleeve and place it back onto the metal table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.2, "hop_absolute_value": 2.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Fold_a_small_tshirt_on_a_wooden_table_with_a_white_background_while_sitting_episode_6-train-bf_000375-af_000050-minus-0000026", "task": "Fold a small tshirt on a wooden table with a white background while sitting", "image": [ "images/human_egodex/Fold_a_small_tshirt_on_a_wooden_table_with_a_white_background_while_sitting/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_wooden_table_with_a_white_background_while_sitting/episode_5/cam_high/frame_000404.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_wooden_table_with_a_white_background_while_sitting/episode_6/cam_high/frame_000375.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_wooden_table_with_a_white_background_while_sitting/episode_6/cam_high/frame_000375.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_wooden_table_with_a_white_background_while_sitting/episode_6/cam_high/frame_000375.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_wooden_table_with_a_white_background_while_sitting/episode_6/cam_high/frame_000050.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_wooden_table_with_a_white_background_while_sitting/episode_6/cam_high/frame_000050.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_wooden_table_with_a_white_background_while_sitting/episode_6/cam_high/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold a small tshirt on a wooden table with a white background while sitting`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold a small tshirt on a wooden table with a white background while sitting`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -86.7, "hop_absolute_value": 86.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unpeel_the_sticker_from_the_stickerbook_and_place_it_back_into_its_numbered_slot_episode_5-train-bf_000050-af_000025-minus-0000020", "task": "Unpeel the sticker from the stickerbook and place it back into its numbered slot", "image": [ "images/human_egodex/Unpeel_the_sticker_from_the_stickerbook_and_place_it_back_into_its_numbered_slot/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Unpeel_the_sticker_from_the_stickerbook_and_place_it_back_into_its_numbered_slot/episode_5/cam_high/frame_000218.jpg", "images/human_egodex/Unpeel_the_sticker_from_the_stickerbook_and_place_it_back_into_its_numbered_slot/episode_5/cam_high/frame_000050.jpg", "images/human_egodex/Unpeel_the_sticker_from_the_stickerbook_and_place_it_back_into_its_numbered_slot/episode_5/cam_high/frame_000050.jpg", "images/human_egodex/Unpeel_the_sticker_from_the_stickerbook_and_place_it_back_into_its_numbered_slot/episode_5/cam_high/frame_000050.jpg", "images/human_egodex/Unpeel_the_sticker_from_the_stickerbook_and_place_it_back_into_its_numbered_slot/episode_5/cam_high/frame_000025.jpg", "images/human_egodex/Unpeel_the_sticker_from_the_stickerbook_and_place_it_back_into_its_numbered_slot/episode_5/cam_high/frame_000025.jpg", "images/human_egodex/Unpeel_the_sticker_from_the_stickerbook_and_place_it_back_into_its_numbered_slot/episode_5/cam_high/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unpeel the sticker from the stickerbook and place it back into its numbered slot`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unpeel the sticker from the stickerbook and place it back into its numbered slot`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate_episode_0-train-bf_000650-af_000500-minus-0000430", "task": "Unwrap the sandwich from the plastic wrap and place it back on the gray plate", "image": [ "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_0/cam_high/frame_000756.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_0/cam_high/frame_000650.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_0/cam_high/frame_000650.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_0/cam_high/frame_000650.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_0/cam_high/frame_000500.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_0/cam_high/frame_000500.jpg", "images/human_egodex/Unwrap_the_sandwich_from_the_plastic_wrap_and_place_it_back_on_the_gray_plate/episode_0/cam_high/frame_000500.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unwrap the sandwich from the plastic wrap and place it back on the gray plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unwrap the sandwich from the plastic wrap and place it back on the gray plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -23.1, "hop_absolute_value": 23.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Open_the_fridge_door_pick_up_three_food_items_from_the_table_place_the_food_items_into_the_fridge_and_close_the_fridge_door_episode_15-train-bf_000350-af_000275-minus-0000241", "task": "Open the fridge door pick up three food items from the table place the food items into the fridge and close the fridge door", "image": [ "images/human_egodex/Open_the_fridge_door_pick_up_three_food_items_from_the_table_place_the_food_items_into_the_fridge_and_close_the_fridge_door/episode_15/cam_high/frame_000000.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_three_food_items_from_the_table_place_the_food_items_into_the_fridge_and_close_the_fridge_door/episode_15/cam_high/frame_000421.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_three_food_items_from_the_table_place_the_food_items_into_the_fridge_and_close_the_fridge_door/episode_15/cam_high/frame_000350.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_three_food_items_from_the_table_place_the_food_items_into_the_fridge_and_close_the_fridge_door/episode_15/cam_high/frame_000350.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_three_food_items_from_the_table_place_the_food_items_into_the_fridge_and_close_the_fridge_door/episode_15/cam_high/frame_000350.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_three_food_items_from_the_table_place_the_food_items_into_the_fridge_and_close_the_fridge_door/episode_15/cam_high/frame_000275.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_three_food_items_from_the_table_place_the_food_items_into_the_fridge_and_close_the_fridge_door/episode_15/cam_high/frame_000275.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_three_food_items_from_the_table_place_the_food_items_into_the_fridge_and_close_the_fridge_door/episode_15/cam_high/frame_000275.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the fridge door pick up three food items from the table place the food items into the fridge and close the fridge door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the fridge door pick up three food items from the table place the food items into the fridge and close the fridge door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -21.4, "hop_absolute_value": 21.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Assemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background_episode_11-train-bf_000850-af_000300-minus-0000085", "task": "Assemble eighteen tiles on a wooden table while sitting against a white background", "image": [ "images/human_egodex/Assemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Assemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_11/cam_high/frame_000914.jpg", "images/human_egodex/Assemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_11/cam_high/frame_000850.jpg", "images/human_egodex/Assemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_11/cam_high/frame_000850.jpg", "images/human_egodex/Assemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_11/cam_high/frame_000850.jpg", "images/human_egodex/Assemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_11/cam_high/frame_000300.jpg", "images/human_egodex/Assemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_11/cam_high/frame_000300.jpg", "images/human_egodex/Assemble_eighteen_tiles_on_a_wooden_table_while_sitting_against_a_white_background/episode_11/cam_high/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Assemble eighteen tiles on a wooden table while sitting against a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Assemble eighteen tiles on a wooden table while sitting against a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -64.7, "hop_absolute_value": 64.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_lid_from_the_tube_remove_the_medium_black_shirt_from_the_tube_and_then_close_the_tube_with_the_lid_episode_14-train-bf_000350-af_000225-minus-0000460", "task": "Remove the lid from the tube remove the medium black shirt from the tube and then close the tube with the lid", "image": [ "images/human_egodex/Remove_the_lid_from_the_tube_remove_the_medium_black_shirt_from_the_tube_and_then_close_the_tube_with_the_lid/episode_12/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_remove_the_medium_black_shirt_from_the_tube_and_then_close_the_tube_with_the_lid/episode_12/cam_high/frame_000674.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_remove_the_medium_black_shirt_from_the_tube_and_then_close_the_tube_with_the_lid/episode_14/cam_high/frame_000350.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_remove_the_medium_black_shirt_from_the_tube_and_then_close_the_tube_with_the_lid/episode_14/cam_high/frame_000350.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_remove_the_medium_black_shirt_from_the_tube_and_then_close_the_tube_with_the_lid/episode_14/cam_high/frame_000350.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_remove_the_medium_black_shirt_from_the_tube_and_then_close_the_tube_with_the_lid/episode_14/cam_high/frame_000225.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_remove_the_medium_black_shirt_from_the_tube_and_then_close_the_tube_with_the_lid/episode_14/cam_high/frame_000225.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_remove_the_medium_black_shirt_from_the_tube_and_then_close_the_tube_with_the_lid/episode_14/cam_high/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the lid from the tube remove the medium black shirt from the tube and then close the tube with the lid`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the lid from the tube remove the medium black shirt from the tube and then close the tube with the lid`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -35.7, "hop_absolute_value": 35.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_Vanguard_card_from_the_card_sleeve_and_place_it_back_on_the_metal_table_episode_1-train-bf_001625-af_001575-minus-0000323", "task": "Remove the Vanguard card from the card sleeve and place it back on the metal table", "image": [ "images/human_egodex/Remove_the_Vanguard_card_from_the_card_sleeve_and_place_it_back_on_the_metal_table/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_card_sleeve_and_place_it_back_on_the_metal_table/episode_2/cam_high/frame_002411.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_card_sleeve_and_place_it_back_on_the_metal_table/episode_1/cam_high/frame_001625.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_card_sleeve_and_place_it_back_on_the_metal_table/episode_1/cam_high/frame_001625.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_card_sleeve_and_place_it_back_on_the_metal_table/episode_1/cam_high/frame_001625.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_card_sleeve_and_place_it_back_on_the_metal_table/episode_1/cam_high/frame_001575.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_card_sleeve_and_place_it_back_on_the_metal_table/episode_1/cam_high/frame_001575.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_card_sleeve_and_place_it_back_on_the_metal_table/episode_1/cam_high/frame_001575.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the Vanguard card from the card sleeve and place it back on the metal table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the Vanguard card from the card sleeve and place it back on the metal table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.1, "hop_absolute_value": 3.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_blue_plushie_from_the_wooden_table_with_the_right_hand_and_place_it_in_the_brown_barrel_episode_10-train-bf_000075-af_000000-minus-0000003", "task": "Pick up the blue plushie from the wooden table with the right hand and place it in the brown barrel", "image": [ "images/human_egodex/Pick_up_the_blue_plushie_from_the_wooden_table_with_the_right_hand_and_place_it_in_the_brown_barrel/episode_10/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_blue_plushie_from_the_wooden_table_with_the_right_hand_and_place_it_in_the_brown_barrel/episode_10/cam_high/frame_000102.jpg", "images/human_egodex/Pick_up_the_blue_plushie_from_the_wooden_table_with_the_right_hand_and_place_it_in_the_brown_barrel/episode_10/cam_high/frame_000075.jpg", "images/human_egodex/Pick_up_the_blue_plushie_from_the_wooden_table_with_the_right_hand_and_place_it_in_the_brown_barrel/episode_10/cam_high/frame_000075.jpg", "images/human_egodex/Pick_up_the_blue_plushie_from_the_wooden_table_with_the_right_hand_and_place_it_in_the_brown_barrel/episode_10/cam_high/frame_000075.jpg", "images/human_egodex/Pick_up_the_blue_plushie_from_the_wooden_table_with_the_right_hand_and_place_it_in_the_brown_barrel/episode_10/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_blue_plushie_from_the_wooden_table_with_the_right_hand_and_place_it_in_the_brown_barrel/episode_10/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_blue_plushie_from_the_wooden_table_with_the_right_hand_and_place_it_in_the_brown_barrel/episode_10/cam_high/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the blue plushie from the wooden table with the right hand and place it in the brown barrel`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the blue plushie from the wooden table with the right hand and place it in the brown barrel`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_a_white_puppy_plush_from_the_blue_bowl_and_place_it_in_the_purple_scoop_episode_13-train-bf_000050-af_000000-minus-0000001", "task": "Pick up a white puppy plush from the blue bowl and place it in the purple scoop", "image": [ "images/human_egodex/Pick_up_a_white_puppy_plush_from_the_blue_bowl_and_place_it_in_the_purple_scoop/episode_13/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_a_white_puppy_plush_from_the_blue_bowl_and_place_it_in_the_purple_scoop/episode_13/cam_high/frame_000134.jpg", "images/human_egodex/Pick_up_a_white_puppy_plush_from_the_blue_bowl_and_place_it_in_the_purple_scoop/episode_13/cam_high/frame_000050.jpg", "images/human_egodex/Pick_up_a_white_puppy_plush_from_the_blue_bowl_and_place_it_in_the_purple_scoop/episode_13/cam_high/frame_000050.jpg", "images/human_egodex/Pick_up_a_white_puppy_plush_from_the_blue_bowl_and_place_it_in_the_purple_scoop/episode_13/cam_high/frame_000050.jpg", "images/human_egodex/Pick_up_a_white_puppy_plush_from_the_blue_bowl_and_place_it_in_the_purple_scoop/episode_13/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_a_white_puppy_plush_from_the_blue_bowl_and_place_it_in_the_purple_scoop/episode_13/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_a_white_puppy_plush_from_the_blue_bowl_and_place_it_in_the_purple_scoop/episode_13/cam_high/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up a white puppy plush from the blue bowl and place it in the purple scoop`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up a white puppy plush from the blue bowl and place it in the purple scoop`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_own_USB_from_white_USB_while_sitting_at_a_green_tablecloth_with_a_blue_background_episode_6-train-bf_000301-af_000250-minus-0000068", "task": "Remove own USB from white USB while sitting at a green tablecloth with a blue background", "image": [ "images/human_egodex/Remove_own_USB_from_white_USB_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Remove_own_USB_from_white_USB_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_18/cam_high/frame_000370.jpg", "images/human_egodex/Remove_own_USB_from_white_USB_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000301.jpg", "images/human_egodex/Remove_own_USB_from_white_USB_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000301.jpg", "images/human_egodex/Remove_own_USB_from_white_USB_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000301.jpg", "images/human_egodex/Remove_own_USB_from_white_USB_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000250.jpg", "images/human_egodex/Remove_own_USB_from_white_USB_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000250.jpg", "images/human_egodex/Remove_own_USB_from_white_USB_while_sitting_at_a_green_tablecloth_with_a_blue_background/episode_6/cam_high/frame_000250.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove own USB from white USB while sitting at a green tablecloth with a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove own USB from white USB while sitting at a green tablecloth with a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_tennis_ball_from_the_container_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background_episode_3-train-bf_000250-af_000100-minus-0000291", "task": "Remove the tennis ball from the container while sitting at a table with a pink tablecloth and lavender background", "image": [ "images/human_egodex/Remove_the_tennis_ball_from_the_container_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_13/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_tennis_ball_from_the_container_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_13/cam_high/frame_000416.jpg", "images/human_egodex/Remove_the_tennis_ball_from_the_container_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000250.jpg", "images/human_egodex/Remove_the_tennis_ball_from_the_container_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000250.jpg", "images/human_egodex/Remove_the_tennis_ball_from_the_container_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000250.jpg", "images/human_egodex/Remove_the_tennis_ball_from_the_container_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000100.jpg", "images/human_egodex/Remove_the_tennis_ball_from_the_container_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000100.jpg", "images/human_egodex/Remove_the_tennis_ball_from_the_container_while_sitting_at_a_table_with_a_pink_tablecloth_and_lavender_background/episode_3/cam_high/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the tennis ball from the container while sitting at a table with a pink tablecloth and lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the tennis ball from the container while sitting at a table with a pink tablecloth and lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Grab_the_sandwich_from_the_gray_plate_and_wrap_it_with_plastic_wrap_while_sitting_against_a_white_background_episode_9-train-bf_000275-af_000200-minus-0000131", "task": "Grab the sandwich from the gray plate and wrap it with plastic wrap while sitting against a white background", "image": [ "images/human_egodex/Grab_the_sandwich_from_the_gray_plate_and_wrap_it_with_plastic_wrap_while_sitting_against_a_white_background/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Grab_the_sandwich_from_the_gray_plate_and_wrap_it_with_plastic_wrap_while_sitting_against_a_white_background/episode_9/cam_high/frame_000813.jpg", "images/human_egodex/Grab_the_sandwich_from_the_gray_plate_and_wrap_it_with_plastic_wrap_while_sitting_against_a_white_background/episode_9/cam_high/frame_000275.jpg", "images/human_egodex/Grab_the_sandwich_from_the_gray_plate_and_wrap_it_with_plastic_wrap_while_sitting_against_a_white_background/episode_9/cam_high/frame_000275.jpg", "images/human_egodex/Grab_the_sandwich_from_the_gray_plate_and_wrap_it_with_plastic_wrap_while_sitting_against_a_white_background/episode_9/cam_high/frame_000275.jpg", "images/human_egodex/Grab_the_sandwich_from_the_gray_plate_and_wrap_it_with_plastic_wrap_while_sitting_against_a_white_background/episode_9/cam_high/frame_000200.jpg", "images/human_egodex/Grab_the_sandwich_from_the_gray_plate_and_wrap_it_with_plastic_wrap_while_sitting_against_a_white_background/episode_9/cam_high/frame_000200.jpg", "images/human_egodex/Grab_the_sandwich_from_the_gray_plate_and_wrap_it_with_plastic_wrap_while_sitting_against_a_white_background/episode_9/cam_high/frame_000200.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab the sandwich from the gray plate and wrap it with plastic wrap while sitting against a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab the sandwich from the gray plate and wrap it with plastic wrap while sitting against a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -27.3, "hop_absolute_value": 27.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Reset_dominoes_upright_on_a_metal_table_with_a_white_background_after_toppling_episode_6-train-bf_001550-af_001075-minus-0000129", "task": "Reset dominoes upright on a metal table with a white background after toppling", "image": [ "images/human_egodex/Reset_dominoes_upright_on_a_metal_table_with_a_white_background_after_toppling/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Reset_dominoes_upright_on_a_metal_table_with_a_white_background_after_toppling/episode_6/cam_high/frame_002251.jpg", "images/human_egodex/Reset_dominoes_upright_on_a_metal_table_with_a_white_background_after_toppling/episode_6/cam_high/frame_001550.jpg", "images/human_egodex/Reset_dominoes_upright_on_a_metal_table_with_a_white_background_after_toppling/episode_6/cam_high/frame_001550.jpg", "images/human_egodex/Reset_dominoes_upright_on_a_metal_table_with_a_white_background_after_toppling/episode_6/cam_high/frame_001550.jpg", "images/human_egodex/Reset_dominoes_upright_on_a_metal_table_with_a_white_background_after_toppling/episode_6/cam_high/frame_001075.jpg", "images/human_egodex/Reset_dominoes_upright_on_a_metal_table_with_a_white_background_after_toppling/episode_6/cam_high/frame_001075.jpg", "images/human_egodex/Reset_dominoes_upright_on_a_metal_table_with_a_white_background_after_toppling/episode_6/cam_high/frame_001075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Reset dominoes upright on a metal table with a white background after toppling`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Reset dominoes upright on a metal table with a white background after toppling`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -30.6, "hop_absolute_value": 30.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background_episode_18-train-bf_000300-af_000125-minus-0000081", "task": "Make a sandwich using breads cheese petty pickles egg lettuce ketchup and mayo on a blue tablecloth with a pink background", "image": [ "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_18/cam_high/frame_000548.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_18/cam_high/frame_000300.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_18/cam_high/frame_000300.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_18/cam_high/frame_000300.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_18/cam_high/frame_000125.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_18/cam_high/frame_000125.jpg", "images/human_egodex/Make_a_sandwich_using_breads_cheese_petty_pickles_egg_lettuce_ketchup_and_mayo_on_a_blue_tablecloth_with_a_pink_background/episode_18/cam_high/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make a sandwich using breads cheese petty pickles egg lettuce ketchup and mayo on a blue tablecloth with a pink background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make a sandwich using breads cheese petty pickles egg lettuce ketchup and mayo on a blue tablecloth with a pink background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -58.3, "hop_absolute_value": 58.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_yellow_highlighter_from_the_orange_bowl_with_the_right_hand_and_place_it_on_the_yellow_cup_episode_10-train-bf_000328-af_000125-minus-0000039", "task": "Pick up the yellow highlighter from the orange bowl with the right hand and place it on the yellow cup", "image": [ "images/human_egodex/Pick_up_the_yellow_highlighter_from_the_orange_bowl_with_the_right_hand_and_place_it_on_the_yellow_cup/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_yellow_highlighter_from_the_orange_bowl_with_the_right_hand_and_place_it_on_the_yellow_cup/episode_7/cam_high/frame_000252.jpg", "images/human_egodex/Pick_up_the_yellow_highlighter_from_the_orange_bowl_with_the_right_hand_and_place_it_on_the_yellow_cup/episode_10/cam_high/frame_000328.jpg", "images/human_egodex/Pick_up_the_yellow_highlighter_from_the_orange_bowl_with_the_right_hand_and_place_it_on_the_yellow_cup/episode_10/cam_high/frame_000328.jpg", "images/human_egodex/Pick_up_the_yellow_highlighter_from_the_orange_bowl_with_the_right_hand_and_place_it_on_the_yellow_cup/episode_10/cam_high/frame_000328.jpg", "images/human_egodex/Pick_up_the_yellow_highlighter_from_the_orange_bowl_with_the_right_hand_and_place_it_on_the_yellow_cup/episode_10/cam_high/frame_000125.jpg", "images/human_egodex/Pick_up_the_yellow_highlighter_from_the_orange_bowl_with_the_right_hand_and_place_it_on_the_yellow_cup/episode_10/cam_high/frame_000125.jpg", "images/human_egodex/Pick_up_the_yellow_highlighter_from_the_orange_bowl_with_the_right_hand_and_place_it_on_the_yellow_cup/episode_10/cam_high/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the yellow highlighter from the orange bowl with the right hand and place it on the yellow cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the yellow highlighter from the orange bowl with the right hand and place it on the yellow cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -61.5, "hop_absolute_value": 61.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_five_legos_on_a_green_tablecloth_while_sitting_against_a_blue_background_episode_6-train-bf_000150-af_000100-minus-0000027", "task": "Disassemble five legos on a green tablecloth while sitting against a blue background", "image": [ "images/human_egodex/Disassemble_five_legos_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_five_legos_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_6/cam_high/frame_000200.jpg", "images/human_egodex/Disassemble_five_legos_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_6/cam_high/frame_000150.jpg", "images/human_egodex/Disassemble_five_legos_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_6/cam_high/frame_000150.jpg", "images/human_egodex/Disassemble_five_legos_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_6/cam_high/frame_000150.jpg", "images/human_egodex/Disassemble_five_legos_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_6/cam_high/frame_000100.jpg", "images/human_egodex/Disassemble_five_legos_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_6/cam_high/frame_000100.jpg", "images/human_egodex/Disassemble_five_legos_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_6/cam_high/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble five legos on a green tablecloth while sitting against a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble five legos on a green tablecloth while sitting against a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_a_vanguard_card_from_a_black_or_red_card_sleeve_and_place_it_back_onto_the_metal_table_episode_19-train-bf_000500-af_000350-minus-0000125", "task": "Remove a vanguard card from a black or red card sleeve and place it back onto the metal table", "image": [ "images/human_egodex/Remove_a_vanguard_card_from_a_black_or_red_card_sleeve_and_place_it_back_onto_the_metal_table/episode_19/cam_high/frame_000000.jpg", "images/human_egodex/Remove_a_vanguard_card_from_a_black_or_red_card_sleeve_and_place_it_back_onto_the_metal_table/episode_19/cam_high/frame_001127.jpg", "images/human_egodex/Remove_a_vanguard_card_from_a_black_or_red_card_sleeve_and_place_it_back_onto_the_metal_table/episode_19/cam_high/frame_000500.jpg", "images/human_egodex/Remove_a_vanguard_card_from_a_black_or_red_card_sleeve_and_place_it_back_onto_the_metal_table/episode_19/cam_high/frame_000500.jpg", "images/human_egodex/Remove_a_vanguard_card_from_a_black_or_red_card_sleeve_and_place_it_back_onto_the_metal_table/episode_19/cam_high/frame_000500.jpg", "images/human_egodex/Remove_a_vanguard_card_from_a_black_or_red_card_sleeve_and_place_it_back_onto_the_metal_table/episode_19/cam_high/frame_000350.jpg", "images/human_egodex/Remove_a_vanguard_card_from_a_black_or_red_card_sleeve_and_place_it_back_onto_the_metal_table/episode_19/cam_high/frame_000350.jpg", "images/human_egodex/Remove_a_vanguard_card_from_a_black_or_red_card_sleeve_and_place_it_back_onto_the_metal_table/episode_19/cam_high/frame_000350.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove a vanguard card from a black or red card sleeve and place it back onto the metal table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove a vanguard card from a black or red card sleeve and place it back onto the metal table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -30.0, "hop_absolute_value": 30.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink_episode_2-train-bf_000825-af_000325-minus-0000079", "task": "Boil water in a pot take an egg from the carton and boil it serve the boiled egg in a gray bowl placed on the round table and place the pot and ladle in the sink", "image": [ "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_18/cam_high/frame_001163.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_2/cam_high/frame_000825.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_2/cam_high/frame_000825.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_2/cam_high/frame_000825.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_2/cam_high/frame_000325.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_2/cam_high/frame_000325.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_carton_and_boil_it_serve_the_boiled_egg_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_2/cam_high/frame_000325.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Boil water in a pot take an egg from the carton and boil it serve the boiled egg in a gray bowl placed on the round table and place the pot and ladle in the sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Boil water in a pot take an egg from the carton and boil it serve the boiled egg in a gray bowl placed on the round table and place the pot and ladle in the sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -60.6, "hop_absolute_value": 60.6, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_lid_from_the_tube_insert_the_green_medium_shirt_into_the_tube_and_then_close_the_tube_with_the_lid_episode_17-train-bf_000125-af_000000-minus-0000003", "task": "Remove the lid from the tube insert the green medium shirt into the tube and then close the tube with the lid", "image": [ "images/human_egodex/Remove_the_lid_from_the_tube_insert_the_green_medium_shirt_into_the_tube_and_then_close_the_tube_with_the_lid/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_insert_the_green_medium_shirt_into_the_tube_and_then_close_the_tube_with_the_lid/episode_7/cam_high/frame_001059.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_insert_the_green_medium_shirt_into_the_tube_and_then_close_the_tube_with_the_lid/episode_17/cam_high/frame_000125.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_insert_the_green_medium_shirt_into_the_tube_and_then_close_the_tube_with_the_lid/episode_17/cam_high/frame_000125.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_insert_the_green_medium_shirt_into_the_tube_and_then_close_the_tube_with_the_lid/episode_17/cam_high/frame_000125.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_insert_the_green_medium_shirt_into_the_tube_and_then_close_the_tube_with_the_lid/episode_17/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_insert_the_green_medium_shirt_into_the_tube_and_then_close_the_tube_with_the_lid/episode_17/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_lid_from_the_tube_insert_the_green_medium_shirt_into_the_tube_and_then_close_the_tube_with_the_lid/episode_17/cam_high/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the lid from the tube insert the green medium shirt into the tube and then close the tube with the lid`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the lid from the tube insert the green medium shirt into the tube and then close the tube with the lid`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_bowl_from_round_table_take_egg_out_of_bowl_return_egg_to_carton_and_empty_pot_of_water_episode_2-train-bf_000675-af_000525-minus-0000412", "task": "Remove bowl from round table take egg out of bowl return egg to carton and empty pot of water", "image": [ "images/human_egodex/Remove_bowl_from_round_table_take_egg_out_of_bowl_return_egg_to_carton_and_empty_pot_of_water/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Remove_bowl_from_round_table_take_egg_out_of_bowl_return_egg_to_carton_and_empty_pot_of_water/episode_2/cam_high/frame_000706.jpg", "images/human_egodex/Remove_bowl_from_round_table_take_egg_out_of_bowl_return_egg_to_carton_and_empty_pot_of_water/episode_2/cam_high/frame_000675.jpg", "images/human_egodex/Remove_bowl_from_round_table_take_egg_out_of_bowl_return_egg_to_carton_and_empty_pot_of_water/episode_2/cam_high/frame_000675.jpg", "images/human_egodex/Remove_bowl_from_round_table_take_egg_out_of_bowl_return_egg_to_carton_and_empty_pot_of_water/episode_2/cam_high/frame_000675.jpg", "images/human_egodex/Remove_bowl_from_round_table_take_egg_out_of_bowl_return_egg_to_carton_and_empty_pot_of_water/episode_2/cam_high/frame_000525.jpg", "images/human_egodex/Remove_bowl_from_round_table_take_egg_out_of_bowl_return_egg_to_carton_and_empty_pot_of_water/episode_2/cam_high/frame_000525.jpg", "images/human_egodex/Remove_bowl_from_round_table_take_egg_out_of_bowl_return_egg_to_carton_and_empty_pot_of_water/episode_2/cam_high/frame_000525.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove bowl from round table take egg out of bowl return egg to carton and empty pot of water`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove bowl from round table take egg out of bowl return egg to carton and empty pot of water`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -22.2, "hop_absolute_value": 22.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_the_assembled_whale_puzzle_on_the_metal_table_episode_3-train-bf_000100-af_000025-minus-0000056", "task": "Disassemble the assembled whale puzzle on the metal table", "image": [ "images/human_egodex/Disassemble_the_assembled_whale_puzzle_on_the_metal_table/episode_3/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_the_assembled_whale_puzzle_on_the_metal_table/episode_3/cam_high/frame_000793.jpg", "images/human_egodex/Disassemble_the_assembled_whale_puzzle_on_the_metal_table/episode_3/cam_high/frame_000100.jpg", "images/human_egodex/Disassemble_the_assembled_whale_puzzle_on_the_metal_table/episode_3/cam_high/frame_000100.jpg", "images/human_egodex/Disassemble_the_assembled_whale_puzzle_on_the_metal_table/episode_3/cam_high/frame_000100.jpg", "images/human_egodex/Disassemble_the_assembled_whale_puzzle_on_the_metal_table/episode_3/cam_high/frame_000025.jpg", "images/human_egodex/Disassemble_the_assembled_whale_puzzle_on_the_metal_table/episode_3/cam_high/frame_000025.jpg", "images/human_egodex/Disassemble_the_assembled_whale_puzzle_on_the_metal_table/episode_3/cam_high/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble the assembled whale puzzle on the metal table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble the assembled whale puzzle on the metal table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Boil_water_in_pot_take_egg_from_carton_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink_episode_1-train-bf_001000-af_000975-minus-0000317", "task": "Boil water in pot take egg from carton boil the egg serve egg in gray bowl on round table and place pot and ladle in sink", "image": [ "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_000000.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_001107.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_001000.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_001000.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_001000.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_000975.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_000975.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_000975.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Boil water in pot take egg from carton boil the egg serve egg in gray bowl on round table and place pot and ladle in sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Boil water in pot take egg from carton boil the egg serve egg in gray bowl on round table and place pot and ladle in sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.5, "hop_absolute_value": 2.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Insert_three_books_into_the_bookshelf_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background_episode_9-train-bf_000350-af_000100-minus-0000051", "task": "Insert three books into the bookshelf while sitting at a table with a blue tablecloth and pink background", "image": [ "images/human_egodex/Insert_three_books_into_the_bookshelf_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Insert_three_books_into_the_bookshelf_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_9/cam_high/frame_000538.jpg", "images/human_egodex/Insert_three_books_into_the_bookshelf_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_9/cam_high/frame_000350.jpg", "images/human_egodex/Insert_three_books_into_the_bookshelf_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_9/cam_high/frame_000350.jpg", "images/human_egodex/Insert_three_books_into_the_bookshelf_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_9/cam_high/frame_000350.jpg", "images/human_egodex/Insert_three_books_into_the_bookshelf_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_9/cam_high/frame_000100.jpg", "images/human_egodex/Insert_three_books_into_the_bookshelf_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_9/cam_high/frame_000100.jpg", "images/human_egodex/Insert_three_books_into_the_bookshelf_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_9/cam_high/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert three books into the bookshelf while sitting at a table with a blue tablecloth and pink background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert three books into the bookshelf while sitting at a table with a blue tablecloth and pink background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Identify_and_remove_objects_including_a_red_shirt_green_shirt_cup_snack_wrapper_and_bottle_from_a_student_desk_with_a_computer_notebook_pens_and_pink_tablecloth_placing_them_into_a_box_episode_8-train-bf_000425-af_000050-minus-0000027", "task": "Identify and remove objects including a red shirt green shirt cup snack wrapper and bottle from a student desk with a computer notebook pens and pink tablecloth placing them into a box", "image": [ "images/human_egodex/Identify_and_remove_objects_including_a_red_shirt_green_shirt_cup_snack_wrapper_and_bottle_from_a_student_desk_with_a_computer_notebook_pens_and_pink_tablecloth_placing_them_into_a_box/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Identify_and_remove_objects_including_a_red_shirt_green_shirt_cup_snack_wrapper_and_bottle_from_a_student_desk_with_a_computer_notebook_pens_and_pink_tablecloth_placing_them_into_a_box/episode_8/cam_high/frame_000563.jpg", "images/human_egodex/Identify_and_remove_objects_including_a_red_shirt_green_shirt_cup_snack_wrapper_and_bottle_from_a_student_desk_with_a_computer_notebook_pens_and_pink_tablecloth_placing_them_into_a_box/episode_8/cam_high/frame_000425.jpg", "images/human_egodex/Identify_and_remove_objects_including_a_red_shirt_green_shirt_cup_snack_wrapper_and_bottle_from_a_student_desk_with_a_computer_notebook_pens_and_pink_tablecloth_placing_them_into_a_box/episode_8/cam_high/frame_000425.jpg", "images/human_egodex/Identify_and_remove_objects_including_a_red_shirt_green_shirt_cup_snack_wrapper_and_bottle_from_a_student_desk_with_a_computer_notebook_pens_and_pink_tablecloth_placing_them_into_a_box/episode_8/cam_high/frame_000425.jpg", "images/human_egodex/Identify_and_remove_objects_including_a_red_shirt_green_shirt_cup_snack_wrapper_and_bottle_from_a_student_desk_with_a_computer_notebook_pens_and_pink_tablecloth_placing_them_into_a_box/episode_8/cam_high/frame_000050.jpg", "images/human_egodex/Identify_and_remove_objects_including_a_red_shirt_green_shirt_cup_snack_wrapper_and_bottle_from_a_student_desk_with_a_computer_notebook_pens_and_pink_tablecloth_placing_them_into_a_box/episode_8/cam_high/frame_000050.jpg", "images/human_egodex/Identify_and_remove_objects_including_a_red_shirt_green_shirt_cup_snack_wrapper_and_bottle_from_a_student_desk_with_a_computer_notebook_pens_and_pink_tablecloth_placing_them_into_a_box/episode_8/cam_high/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Identify and remove objects including a red shirt green shirt cup snack wrapper and bottle from a student desk with a computer notebook pens and pink tablecloth placing them into a box`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Identify and remove objects including a red shirt green shirt cup snack wrapper and bottle from a student desk with a computer notebook pens and pink tablecloth placing them into a box`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -88.2, "hop_absolute_value": 88.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Insert_blacktip_AirPods_while_sitting_at_a_wooden_table_with_a_brown_background_episode_2-train-bf_000125-af_000075-minus-0000019", "task": "Insert blacktip AirPods while sitting at a wooden table with a brown background", "image": [ "images/human_egodex/Insert_blacktip_AirPods_while_sitting_at_a_wooden_table_with_a_brown_background/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Insert_blacktip_AirPods_while_sitting_at_a_wooden_table_with_a_brown_background/episode_9/cam_high/frame_000173.jpg", "images/human_egodex/Insert_blacktip_AirPods_while_sitting_at_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000125.jpg", "images/human_egodex/Insert_blacktip_AirPods_while_sitting_at_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000125.jpg", "images/human_egodex/Insert_blacktip_AirPods_while_sitting_at_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000125.jpg", "images/human_egodex/Insert_blacktip_AirPods_while_sitting_at_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000075.jpg", "images/human_egodex/Insert_blacktip_AirPods_while_sitting_at_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000075.jpg", "images/human_egodex/Insert_blacktip_AirPods_while_sitting_at_a_wooden_table_with_a_brown_background/episode_2/cam_high/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert blacktip AirPods while sitting at a wooden table with a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert blacktip AirPods while sitting at a wooden table with a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_puzzle2_episode_0-train-bf_001150-af_000725-minus-0000117", "task": "Pick up the puzzle pieces from a wooden table with the right hand and assemble puzzle2", "image": [ "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_puzzle2/episode_17/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_puzzle2/episode_17/cam_high/frame_001609.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_puzzle2/episode_0/cam_high/frame_001150.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_puzzle2/episode_0/cam_high/frame_001150.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_puzzle2/episode_0/cam_high/frame_001150.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_puzzle2/episode_0/cam_high/frame_000725.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_puzzle2/episode_0/cam_high/frame_000725.jpg", "images/human_egodex/Pick_up_the_puzzle_pieces_from_a_wooden_table_with_the_right_hand_and_assemble_puzzle2/episode_0/cam_high/frame_000725.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the puzzle pieces from a wooden table with the right hand and assemble puzzle2`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the puzzle pieces from a wooden table with the right hand and assemble puzzle2`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -37.0, "hop_absolute_value": 37.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_lid_from_the_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background_episode_18-train-bf_000300-af_000075-minus-0000039", "task": "Remove the lid from the cup placed on a wooden table while sitting against a brown background", "image": [ "images/human_egodex/Remove_the_lid_from_the_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_lid_from_the_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_18/cam_high/frame_000392.jpg", "images/human_egodex/Remove_the_lid_from_the_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_18/cam_high/frame_000300.jpg", "images/human_egodex/Remove_the_lid_from_the_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_18/cam_high/frame_000300.jpg", "images/human_egodex/Remove_the_lid_from_the_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_18/cam_high/frame_000300.jpg", "images/human_egodex/Remove_the_lid_from_the_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_18/cam_high/frame_000075.jpg", "images/human_egodex/Remove_the_lid_from_the_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_18/cam_high/frame_000075.jpg", "images/human_egodex/Remove_the_lid_from_the_cup_placed_on_a_wooden_table_while_sitting_against_a_brown_background/episode_18/cam_high/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the lid from the cup placed on a wooden table while sitting against a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the lid from the cup placed on a wooden table while sitting against a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Assemble_the_seal_puzzle_on_a_metal_table_while_sitting_with_a_white_background_episode_7-train-bf_001500-af_000925-minus-0000119", "task": "Assemble the seal puzzle on a metal table while sitting with a white background", "image": [ "images/human_egodex/Assemble_the_seal_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Assemble_the_seal_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_7/cam_high/frame_002435.jpg", "images/human_egodex/Assemble_the_seal_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_7/cam_high/frame_001500.jpg", "images/human_egodex/Assemble_the_seal_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_7/cam_high/frame_001500.jpg", "images/human_egodex/Assemble_the_seal_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_7/cam_high/frame_001500.jpg", "images/human_egodex/Assemble_the_seal_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_7/cam_high/frame_000925.jpg", "images/human_egodex/Assemble_the_seal_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_7/cam_high/frame_000925.jpg", "images/human_egodex/Assemble_the_seal_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_7/cam_high/frame_000925.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Assemble the seal puzzle on a metal table while sitting with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Assemble the seal puzzle on a metal table while sitting with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -38.3, "hop_absolute_value": 38.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unplug_the_lightning_cable_from_the_small_white_phone_placed_on_the_green_tablecloth_episode_15-train-bf_000050-af_000000-minus-0000001", "task": "Unplug the lightning cable from the small white phone placed on the green tablecloth", "image": [ "images/human_egodex/Unplug_the_lightning_cable_from_the_small_white_phone_placed_on_the_green_tablecloth/episode_15/cam_high/frame_000000.jpg", "images/human_egodex/Unplug_the_lightning_cable_from_the_small_white_phone_placed_on_the_green_tablecloth/episode_15/cam_high/frame_000139.jpg", "images/human_egodex/Unplug_the_lightning_cable_from_the_small_white_phone_placed_on_the_green_tablecloth/episode_15/cam_high/frame_000050.jpg", "images/human_egodex/Unplug_the_lightning_cable_from_the_small_white_phone_placed_on_the_green_tablecloth/episode_15/cam_high/frame_000050.jpg", "images/human_egodex/Unplug_the_lightning_cable_from_the_small_white_phone_placed_on_the_green_tablecloth/episode_15/cam_high/frame_000050.jpg", "images/human_egodex/Unplug_the_lightning_cable_from_the_small_white_phone_placed_on_the_green_tablecloth/episode_15/cam_high/frame_000000.jpg", "images/human_egodex/Unplug_the_lightning_cable_from_the_small_white_phone_placed_on_the_green_tablecloth/episode_15/cam_high/frame_000000.jpg", "images/human_egodex/Unplug_the_lightning_cable_from_the_small_white_phone_placed_on_the_green_tablecloth/episode_15/cam_high/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unplug the lightning cable from the small white phone placed on the green tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unplug the lightning cable from the small white phone placed on the green tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_Vanguard_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table_episode_0-train-bf_000475-af_000000-minus-0000007", "task": "Remove the Vanguard card from the black card sleeve and place it back on the metal table", "image": [ "images/human_egodex/Remove_the_Vanguard_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_0/cam_high/frame_001438.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_0/cam_high/frame_000475.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_0/cam_high/frame_000475.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_0/cam_high/frame_000475.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_Vanguard_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_0/cam_high/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the Vanguard card from the black card sleeve and place it back on the metal table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the Vanguard card from the black card sleeve and place it back on the metal table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Make_a_sandwich_using_breads_eggs_bacon_onion_tomato_cheese_and_ketchup_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth_episode_11-train-bf_000150-af_000025-minus-0000034", "task": "Make a sandwich using breads eggs bacon onion tomato cheese and ketchup while sitting at a table with a blue background and no tablecloth", "image": [ "images/human_egodex/Make_a_sandwich_using_breads_eggs_bacon_onion_tomato_cheese_and_ketchup_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Make_a_sandwich_using_breads_eggs_bacon_onion_tomato_cheese_and_ketchup_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_11/cam_high/frame_000510.jpg", "images/human_egodex/Make_a_sandwich_using_breads_eggs_bacon_onion_tomato_cheese_and_ketchup_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_11/cam_high/frame_000150.jpg", "images/human_egodex/Make_a_sandwich_using_breads_eggs_bacon_onion_tomato_cheese_and_ketchup_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_11/cam_high/frame_000150.jpg", "images/human_egodex/Make_a_sandwich_using_breads_eggs_bacon_onion_tomato_cheese_and_ketchup_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_11/cam_high/frame_000150.jpg", "images/human_egodex/Make_a_sandwich_using_breads_eggs_bacon_onion_tomato_cheese_and_ketchup_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_11/cam_high/frame_000025.jpg", "images/human_egodex/Make_a_sandwich_using_breads_eggs_bacon_onion_tomato_cheese_and_ketchup_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_11/cam_high/frame_000025.jpg", "images/human_egodex/Make_a_sandwich_using_breads_eggs_bacon_onion_tomato_cheese_and_ketchup_while_sitting_at_a_table_with_a_blue_background_and_no_tablecloth/episode_11/cam_high/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make a sandwich using breads eggs bacon onion tomato cheese and ketchup while sitting at a table with a blue background and no tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make a sandwich using breads eggs bacon onion tomato cheese and ketchup while sitting at a table with a blue background and no tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Boil_water_in_pot_take_egg_from_carton_and_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink_episode_1-train-bf_000950-af_000025-minus-0000016", "task": "Boil water in pot take egg from carton and boil the egg serve egg in gray bowl on round table and place pot and ladle in sink", "image": [ "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_and_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_12/cam_high/frame_000000.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_and_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_12/cam_high/frame_000427.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_and_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_000950.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_and_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_000950.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_and_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_000950.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_and_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_000025.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_and_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_000025.jpg", "images/human_egodex/Boil_water_in_pot_take_egg_from_carton_and_boil_the_egg_serve_egg_in_gray_bowl_on_round_table_and_place_pot_and_ladle_in_sink/episode_1/cam_high/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Boil water in pot take egg from carton and boil the egg serve egg in gray bowl on round table and place pot and ladle in sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Boil water in pot take egg from carton and boil the egg serve egg in gray bowl on round table and place pot and ladle in sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -97.4, "hop_absolute_value": 97.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Open_the_fridge_door_pick_up_seven_food_items_from_the_table_place_the_food_into_the_fridge_and_close_the_fridge_door_episode_16-train-bf_000675-af_000275-minus-0000093", "task": "Open the fridge door pick up seven food items from the table place the food into the fridge and close the fridge door", "image": [ "images/human_egodex/Open_the_fridge_door_pick_up_seven_food_items_from_the_table_place_the_food_into_the_fridge_and_close_the_fridge_door/episode_16/cam_high/frame_000000.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_seven_food_items_from_the_table_place_the_food_into_the_fridge_and_close_the_fridge_door/episode_16/cam_high/frame_000710.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_seven_food_items_from_the_table_place_the_food_into_the_fridge_and_close_the_fridge_door/episode_16/cam_high/frame_000675.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_seven_food_items_from_the_table_place_the_food_into_the_fridge_and_close_the_fridge_door/episode_16/cam_high/frame_000675.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_seven_food_items_from_the_table_place_the_food_into_the_fridge_and_close_the_fridge_door/episode_16/cam_high/frame_000675.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_seven_food_items_from_the_table_place_the_food_into_the_fridge_and_close_the_fridge_door/episode_16/cam_high/frame_000275.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_seven_food_items_from_the_table_place_the_food_into_the_fridge_and_close_the_fridge_door/episode_16/cam_high/frame_000275.jpg", "images/human_egodex/Open_the_fridge_door_pick_up_seven_food_items_from_the_table_place_the_food_into_the_fridge_and_close_the_fridge_door/episode_16/cam_high/frame_000275.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the fridge door pick up seven food items from the table place the food into the fridge and close the fridge door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the fridge door pick up seven food items from the table place the food into the fridge and close the fridge door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -59.3, "hop_absolute_value": 59.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Braid_the_three_strands_of_salmonpink_yarn_episode_15-train-bf_000325-af_000100-minus-0000042", "task": "Braid the three strands of salmonpink yarn", "image": [ "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_15/cam_high/frame_000000.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_15/cam_high/frame_000386.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_15/cam_high/frame_000325.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_15/cam_high/frame_000325.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_15/cam_high/frame_000325.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_15/cam_high/frame_000100.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_15/cam_high/frame_000100.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn/episode_15/cam_high/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Braid the three strands of salmonpink yarn`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Braid the three strands of salmonpink yarn`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -69.2, "hop_absolute_value": 69.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Fold_a_small_tshirt_on_a_green_tablecloth_while_sitting_against_a_blue_background_episode_8-train-bf_000375-af_000325-minus-0000286", "task": "Fold a small tshirt on a green tablecloth while sitting against a blue background", "image": [ "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_8/cam_high/frame_000516.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_8/cam_high/frame_000375.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_8/cam_high/frame_000375.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_8/cam_high/frame_000375.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_8/cam_high/frame_000325.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_8/cam_high/frame_000325.jpg", "images/human_egodex/Fold_a_small_tshirt_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_8/cam_high/frame_000325.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fold a small tshirt on a green tablecloth while sitting against a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fold a small tshirt on a green tablecloth while sitting against a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -13.3, "hop_absolute_value": 13.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_the_structure_composed_of_11_objects_while_sitting_at_a_table_with_a_lavender_background_episode_15-train-bf_000225-af_000050-minus-0000030", "task": "Disassemble the structure composed of 11 objects while sitting at a table with a lavender background", "image": [ "images/human_egodex/Disassemble_the_structure_composed_of_11_objects_while_sitting_at_a_table_with_a_lavender_background/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_the_structure_composed_of_11_objects_while_sitting_at_a_table_with_a_lavender_background/episode_11/cam_high/frame_000591.jpg", "images/human_egodex/Disassemble_the_structure_composed_of_11_objects_while_sitting_at_a_table_with_a_lavender_background/episode_15/cam_high/frame_000225.jpg", "images/human_egodex/Disassemble_the_structure_composed_of_11_objects_while_sitting_at_a_table_with_a_lavender_background/episode_15/cam_high/frame_000225.jpg", "images/human_egodex/Disassemble_the_structure_composed_of_11_objects_while_sitting_at_a_table_with_a_lavender_background/episode_15/cam_high/frame_000225.jpg", "images/human_egodex/Disassemble_the_structure_composed_of_11_objects_while_sitting_at_a_table_with_a_lavender_background/episode_15/cam_high/frame_000050.jpg", "images/human_egodex/Disassemble_the_structure_composed_of_11_objects_while_sitting_at_a_table_with_a_lavender_background/episode_15/cam_high/frame_000050.jpg", "images/human_egodex/Disassemble_the_structure_composed_of_11_objects_while_sitting_at_a_table_with_a_lavender_background/episode_15/cam_high/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble the structure composed of 11 objects while sitting at a table with a lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble the structure composed of 11 objects while sitting at a table with a lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -77.8, "hop_absolute_value": 77.8, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background_episode_19-train-bf_000375-af_000300-minus-0000128", "task": "Disassemble the lobster puzzle on a metal table while sitting with a white background", "image": [ "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_19/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_19/cam_high/frame_000571.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_19/cam_high/frame_000375.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_19/cam_high/frame_000375.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_19/cam_high/frame_000375.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_19/cam_high/frame_000300.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_19/cam_high/frame_000300.jpg", "images/human_egodex/Disassemble_the_lobster_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_19/cam_high/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble the lobster puzzle on a metal table while sitting with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble the lobster puzzle on a metal table while sitting with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_held_in_hand_and_place_the_stapled_paper_into_a_new_stack_episode_17-train-bf_000600-af_000575-minus-0000378", "task": "Pick up one piece of paper from a stack staple it with a stapler held in hand and place the stapled paper into a new stack", "image": [ "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_held_in_hand_and_place_the_stapled_paper_into_a_new_stack/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_held_in_hand_and_place_the_stapled_paper_into_a_new_stack/episode_7/cam_high/frame_002117.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_held_in_hand_and_place_the_stapled_paper_into_a_new_stack/episode_17/cam_high/frame_000600.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_held_in_hand_and_place_the_stapled_paper_into_a_new_stack/episode_17/cam_high/frame_000600.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_held_in_hand_and_place_the_stapled_paper_into_a_new_stack/episode_17/cam_high/frame_000600.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_held_in_hand_and_place_the_stapled_paper_into_a_new_stack/episode_17/cam_high/frame_000575.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_held_in_hand_and_place_the_stapled_paper_into_a_new_stack/episode_17/cam_high/frame_000575.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_held_in_hand_and_place_the_stapled_paper_into_a_new_stack/episode_17/cam_high/frame_000575.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up one piece of paper from a stack staple it with a stapler held in hand and place the stapled paper into a new stack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up one piece of paper from a stack staple it with a stapler held in hand and place the stapled paper into a new stack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.2, "hop_absolute_value": 4.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Use_chopsticks_to_move_cubes_irregular_shaped_cubes_and_ribbons_into_bowls_while_sitting_at_a_wood_tablecloth_with_a_white_background_episode_0-train-bf_000325-af_000300-minus-0000423", "task": "Use chopsticks to move cubes irregular shaped cubes and ribbons into bowls while sitting at a wood tablecloth with a white background", "image": [ "images/human_egodex/Use_chopsticks_to_move_cubes_irregular_shaped_cubes_and_ribbons_into_bowls_while_sitting_at_a_wood_tablecloth_with_a_white_background/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_irregular_shaped_cubes_and_ribbons_into_bowls_while_sitting_at_a_wood_tablecloth_with_a_white_background/episode_0/cam_high/frame_001385.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_irregular_shaped_cubes_and_ribbons_into_bowls_while_sitting_at_a_wood_tablecloth_with_a_white_background/episode_0/cam_high/frame_000325.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_irregular_shaped_cubes_and_ribbons_into_bowls_while_sitting_at_a_wood_tablecloth_with_a_white_background/episode_0/cam_high/frame_000325.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_irregular_shaped_cubes_and_ribbons_into_bowls_while_sitting_at_a_wood_tablecloth_with_a_white_background/episode_0/cam_high/frame_000325.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_irregular_shaped_cubes_and_ribbons_into_bowls_while_sitting_at_a_wood_tablecloth_with_a_white_background/episode_0/cam_high/frame_000300.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_irregular_shaped_cubes_and_ribbons_into_bowls_while_sitting_at_a_wood_tablecloth_with_a_white_background/episode_0/cam_high/frame_000300.jpg", "images/human_egodex/Use_chopsticks_to_move_cubes_irregular_shaped_cubes_and_ribbons_into_bowls_while_sitting_at_a_wood_tablecloth_with_a_white_background/episode_0/cam_high/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use chopsticks to move cubes irregular shaped cubes and ribbons into bowls while sitting at a wood tablecloth with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use chopsticks to move cubes irregular shaped cubes and ribbons into bowls while sitting at a wood tablecloth with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.7, "hop_absolute_value": 7.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table_episode_13-train-bf_000525-af_000500-minus-0000410", "task": "Remove the Pokemon card from the black card sleeve and place it back on the metal table", "image": [ "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_13/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_13/cam_high/frame_001564.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_13/cam_high/frame_000525.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_13/cam_high/frame_000525.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_13/cam_high/frame_000525.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_13/cam_high/frame_000500.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_13/cam_high/frame_000500.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_13/cam_high/frame_000500.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the Pokemon card from the black card sleeve and place it back on the metal table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the Pokemon card from the black card sleeve and place it back on the metal table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.8, "hop_absolute_value": 4.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Assemble_the_whale_puzzle_on_a_metal_table_while_sitting_with_a_white_background_episode_8-train-bf_001050-af_000525-minus-0000109", "task": "Assemble the whale puzzle on a metal table while sitting with a white background", "image": [ "images/human_egodex/Assemble_the_whale_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Assemble_the_whale_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_8/cam_high/frame_004072.jpg", "images/human_egodex/Assemble_the_whale_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_8/cam_high/frame_001050.jpg", "images/human_egodex/Assemble_the_whale_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_8/cam_high/frame_001050.jpg", "images/human_egodex/Assemble_the_whale_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_8/cam_high/frame_001050.jpg", "images/human_egodex/Assemble_the_whale_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_8/cam_high/frame_000525.jpg", "images/human_egodex/Assemble_the_whale_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_8/cam_high/frame_000525.jpg", "images/human_egodex/Assemble_the_whale_puzzle_on_a_metal_table_while_sitting_with_a_white_background/episode_8/cam_high/frame_000525.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Assemble the whale puzzle on a metal table while sitting with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Assemble the whale puzzle on a metal table while sitting with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Crumple_a_paper_placed_on_a_pink_tablecloth_with_a_lavender_background_episode_16-train-bf_000150-af_000125-minus-0000043", "task": "Crumple a paper placed on a pink tablecloth with a lavender background", "image": [ "images/human_egodex/Crumple_a_paper_placed_on_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000000.jpg", "images/human_egodex/Crumple_a_paper_placed_on_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000231.jpg", "images/human_egodex/Crumple_a_paper_placed_on_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000150.jpg", "images/human_egodex/Crumple_a_paper_placed_on_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000150.jpg", "images/human_egodex/Crumple_a_paper_placed_on_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000150.jpg", "images/human_egodex/Crumple_a_paper_placed_on_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000125.jpg", "images/human_egodex/Crumple_a_paper_placed_on_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000125.jpg", "images/human_egodex/Crumple_a_paper_placed_on_a_pink_tablecloth_with_a_lavender_background/episode_16/cam_high/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Crumple a paper placed on a pink tablecloth with a lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Crumple a paper placed on a pink tablecloth with a lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background_episode_7-train-bf_000325-af_000275-minus-0000289", "task": "Pick up the measuring tape and measure the dimensions of the medium box while sitting at a table with a white background", "image": [ "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_7/cam_high/frame_000529.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_7/cam_high/frame_000325.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_7/cam_high/frame_000325.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_7/cam_high/frame_000325.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_7/cam_high/frame_000275.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_7/cam_high/frame_000275.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_7/cam_high/frame_000275.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the measuring tape and measure the dimensions of the medium box while sitting at a table with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the measuring tape and measure the dimensions of the medium box while sitting at a table with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -15.4, "hop_absolute_value": 15.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_individual_blocks_and_form_3D_figures_on_a_board_placed_on_a_red_tablecloth_episode_10-train-bf_000975-af_000950-minus-0000431", "task": "Pick up individual blocks and form 3D figures on a board placed on a red tablecloth", "image": [ "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_placed_on_a_red_tablecloth/episode_12/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_placed_on_a_red_tablecloth/episode_12/cam_high/frame_001258.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_placed_on_a_red_tablecloth/episode_10/cam_high/frame_000975.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_placed_on_a_red_tablecloth/episode_10/cam_high/frame_000975.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_placed_on_a_red_tablecloth/episode_10/cam_high/frame_000975.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_placed_on_a_red_tablecloth/episode_10/cam_high/frame_000950.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_placed_on_a_red_tablecloth/episode_10/cam_high/frame_000950.jpg", "images/human_egodex/Pick_up_individual_blocks_and_form_3D_figures_on_a_board_placed_on_a_red_tablecloth/episode_10/cam_high/frame_000950.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up individual blocks and form 3D figures on a board placed on a red tablecloth`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up individual blocks and form 3D figures on a board placed on a red tablecloth`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.6, "hop_absolute_value": 2.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_USB_from_the_hairbrush_using_the_right_hand_while_sitting_with_a_blue_background_episode_15-train-bf_000250-af_000225-minus-0000463", "task": "Remove the USB from the hairbrush using the right hand while sitting with a blue background", "image": [ "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_right_hand_while_sitting_with_a_blue_background/episode_16/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_right_hand_while_sitting_with_a_blue_background/episode_16/cam_high/frame_000654.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000250.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000250.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000250.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000225.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000225.jpg", "images/human_egodex/Remove_the_USB_from_the_hairbrush_using_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the USB from the hairbrush using the right hand while sitting with a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the USB from the hairbrush using the right hand while sitting with a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table_episode_16-train-bf_001125-af_000075-minus-0000012", "task": "Remove the Pokemon card from the black card sleeve and place it back on the metal table", "image": [ "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_16/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_16/cam_high/frame_001495.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_16/cam_high/frame_001125.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_16/cam_high/frame_001125.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_16/cam_high/frame_001125.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_16/cam_high/frame_000075.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_16/cam_high/frame_000075.jpg", "images/human_egodex/Remove_the_Pokemon_card_from_the_black_card_sleeve_and_place_it_back_on_the_metal_table/episode_16/cam_high/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the Pokemon card from the black card sleeve and place it back on the metal table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the Pokemon card from the black card sleeve and place it back on the metal table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -93.3, "hop_absolute_value": 93.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Braid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background_episode_5-train-bf_002050-af_002000-minus-0000349", "task": "Braid the strands of the fishtail yarn while sitting against a brown background", "image": [ "images/human_egodex/Braid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Braid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_5/cam_high/frame_002399.jpg", "images/human_egodex/Braid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_5/cam_high/frame_002050.jpg", "images/human_egodex/Braid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_5/cam_high/frame_002050.jpg", "images/human_egodex/Braid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_5/cam_high/frame_002050.jpg", "images/human_egodex/Braid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_5/cam_high/frame_002000.jpg", "images/human_egodex/Braid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_5/cam_high/frame_002000.jpg", "images/human_egodex/Braid_the_strands_of_the_fishtail_yarn_while_sitting_against_a_brown_background/episode_5/cam_high/frame_002000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Braid the strands of the fishtail yarn while sitting against a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Braid the strands of the fishtail yarn while sitting against a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.4, "hop_absolute_value": 2.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background_episode_12-train-bf_001100-af_001025-minus-0000429", "task": "Insert a USB into the hairbrush with the right hand while sitting against a blue background", "image": [ "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_2/cam_high/frame_000000.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_2/cam_high/frame_000716.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_001100.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_001100.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_001100.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_001025.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_001025.jpg", "images/human_egodex/Insert_a_USB_into_the_hairbrush_with_the_right_hand_while_sitting_against_a_blue_background/episode_12/cam_high/frame_001025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert a USB into the hairbrush with the right hand while sitting against a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert a USB into the hairbrush with the right hand while sitting against a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.8, "hop_absolute_value": 6.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Screw_the_screw_into_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background_episode_1-train-bf_000479-af_000300-minus-0000098", "task": "Screw the screw into the fixture while sitting at a table with a blue tablecloth and pink background", "image": [ "images/human_egodex/Screw_the_screw_into_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_1/cam_high/frame_000000.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_1/cam_high/frame_000479.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_1/cam_high/frame_000479.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_1/cam_high/frame_000479.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_1/cam_high/frame_000479.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_1/cam_high/frame_000300.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_1/cam_high/frame_000300.jpg", "images/human_egodex/Screw_the_screw_into_the_fixture_while_sitting_at_a_table_with_a_blue_tablecloth_and_pink_background/episode_1/cam_high/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Screw the screw into the fixture while sitting at a table with a blue tablecloth and pink background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Screw the screw into the fixture while sitting at a table with a blue tablecloth and pink background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -36.8, "hop_absolute_value": 36.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background_episode_6-train-bf_000446-af_000325-minus-0000261", "task": "Pick up the measuring tape and measure the dimensions of the medium box while sitting at a table with a white background", "image": [ "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_4/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_4/cam_high/frame_000660.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_6/cam_high/frame_000446.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_6/cam_high/frame_000446.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_6/cam_high/frame_000446.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_6/cam_high/frame_000325.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_6/cam_high/frame_000325.jpg", "images/human_egodex/Pick_up_the_measuring_tape_and_measure_the_dimensions_of_the_medium_box_while_sitting_at_a_table_with_a_white_background/episode_6/cam_high/frame_000325.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the measuring tape and measure the dimensions of the medium box while sitting at a table with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the measuring tape and measure the dimensions of the medium box while sitting at a table with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -23.5, "hop_absolute_value": 23.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Boil_water_in_a_pot_take_an_egg_from_the_egg_carton_boil_the_egg_serve_it_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink_episode_8-train-bf_001025-af_001000-minus-0000300", "task": "Boil water in a pot take an egg from the egg carton boil the egg serve it in a gray bowl placed on the round table and place the pot and ladle in the sink", "image": [ "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_egg_carton_boil_the_egg_serve_it_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_egg_carton_boil_the_egg_serve_it_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_001063.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_egg_carton_boil_the_egg_serve_it_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_001025.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_egg_carton_boil_the_egg_serve_it_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_001025.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_egg_carton_boil_the_egg_serve_it_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_001025.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_egg_carton_boil_the_egg_serve_it_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_001000.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_egg_carton_boil_the_egg_serve_it_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_001000.jpg", "images/human_egodex/Boil_water_in_a_pot_take_an_egg_from_the_egg_carton_boil_the_egg_serve_it_in_a_gray_bowl_placed_on_the_round_table_and_place_the_pot_and_ladle_in_the_sink/episode_8/cam_high/frame_001000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Boil water in a pot take an egg from the egg carton boil the egg serve it in a gray bowl placed on the round table and place the pot and ladle in the sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Boil water in a pot take an egg from the egg carton boil the egg serve it in a gray bowl placed on the round table and place the pot and ladle in the sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.4, "hop_absolute_value": 2.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_vanguard_card_from_the_card_sleeve_and_place_it_back_onto_the_metal_table_episode_3-train-bf_003175-af_003150-minus-0000333", "task": "Remove the vanguard card from the card sleeve and place it back onto the metal table", "image": [ "images/human_egodex/Remove_the_vanguard_card_from_the_card_sleeve_and_place_it_back_onto_the_metal_table/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_vanguard_card_from_the_card_sleeve_and_place_it_back_onto_the_metal_table/episode_9/cam_high/frame_001809.jpg", "images/human_egodex/Remove_the_vanguard_card_from_the_card_sleeve_and_place_it_back_onto_the_metal_table/episode_3/cam_high/frame_003175.jpg", "images/human_egodex/Remove_the_vanguard_card_from_the_card_sleeve_and_place_it_back_onto_the_metal_table/episode_3/cam_high/frame_003175.jpg", "images/human_egodex/Remove_the_vanguard_card_from_the_card_sleeve_and_place_it_back_onto_the_metal_table/episode_3/cam_high/frame_003175.jpg", "images/human_egodex/Remove_the_vanguard_card_from_the_card_sleeve_and_place_it_back_onto_the_metal_table/episode_3/cam_high/frame_003150.jpg", "images/human_egodex/Remove_the_vanguard_card_from_the_card_sleeve_and_place_it_back_onto_the_metal_table/episode_3/cam_high/frame_003150.jpg", "images/human_egodex/Remove_the_vanguard_card_from_the_card_sleeve_and_place_it_back_onto_the_metal_table/episode_3/cam_high/frame_003150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the vanguard card from the card sleeve and place it back onto the metal table`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the vanguard card from the card sleeve and place it back onto the metal table`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -0.8, "hop_absolute_value": 0.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unfold_the_red_tablecloth_and_set_up_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_on_the_tablecloth_while_sitting_with_a_white_background_episode_17-train-bf_001350-af_001300-minus-0000329", "task": "Unfold the red tablecloth and set up the blue and yellow plates forks spoons knives and cups on the tablecloth while sitting with a white background", "image": [ "images/human_egodex/Unfold_the_red_tablecloth_and_set_up_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_on_the_tablecloth_while_sitting_with_a_white_background/episode_18/cam_high/frame_000000.jpg", "images/human_egodex/Unfold_the_red_tablecloth_and_set_up_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_on_the_tablecloth_while_sitting_with_a_white_background/episode_18/cam_high/frame_001668.jpg", "images/human_egodex/Unfold_the_red_tablecloth_and_set_up_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_on_the_tablecloth_while_sitting_with_a_white_background/episode_17/cam_high/frame_001350.jpg", "images/human_egodex/Unfold_the_red_tablecloth_and_set_up_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_on_the_tablecloth_while_sitting_with_a_white_background/episode_17/cam_high/frame_001350.jpg", "images/human_egodex/Unfold_the_red_tablecloth_and_set_up_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_on_the_tablecloth_while_sitting_with_a_white_background/episode_17/cam_high/frame_001350.jpg", "images/human_egodex/Unfold_the_red_tablecloth_and_set_up_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_on_the_tablecloth_while_sitting_with_a_white_background/episode_17/cam_high/frame_001300.jpg", "images/human_egodex/Unfold_the_red_tablecloth_and_set_up_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_on_the_tablecloth_while_sitting_with_a_white_background/episode_17/cam_high/frame_001300.jpg", "images/human_egodex/Unfold_the_red_tablecloth_and_set_up_the_blue_and_yellow_plates_forks_spoons_knives_and_cups_on_the_tablecloth_while_sitting_with_a_white_background/episode_17/cam_high/frame_001300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unfold the red tablecloth and set up the blue and yellow plates forks spoons knives and cups on the tablecloth while sitting with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unfold the red tablecloth and set up the blue and yellow plates forks spoons knives and cups on the tablecloth while sitting with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.7, "hop_absolute_value": 3.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Unstack_three_bowls_on_a_green_tablecloth_while_sitting_against_a_blue_background_episode_1-train-bf_000150-af_000125-minus-0000174", "task": "Unstack three bowls on a green tablecloth while sitting against a blue background", "image": [ "images/human_egodex/Unstack_three_bowls_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_1/cam_high/frame_000000.jpg", "images/human_egodex/Unstack_three_bowls_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_1/cam_high/frame_000342.jpg", "images/human_egodex/Unstack_three_bowls_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_1/cam_high/frame_000150.jpg", "images/human_egodex/Unstack_three_bowls_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_1/cam_high/frame_000150.jpg", "images/human_egodex/Unstack_three_bowls_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_1/cam_high/frame_000150.jpg", "images/human_egodex/Unstack_three_bowls_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_1/cam_high/frame_000125.jpg", "images/human_egodex/Unstack_three_bowls_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_1/cam_high/frame_000125.jpg", "images/human_egodex/Unstack_three_bowls_on_a_green_tablecloth_while_sitting_against_a_blue_background/episode_1/cam_high/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Unstack three bowls on a green tablecloth while sitting against a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Unstack three bowls on a green tablecloth while sitting against a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_and_place_the_stapled_paper_into_a_new_stack_episode_8-train-bf_000675-af_000650-minus-0000389", "task": "Pick up one piece of paper from a stack staple it with a stapler and place the stapled paper into a new stack", "image": [ "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_and_place_the_stapled_paper_into_a_new_stack/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_and_place_the_stapled_paper_into_a_new_stack/episode_8/cam_high/frame_002007.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_and_place_the_stapled_paper_into_a_new_stack/episode_8/cam_high/frame_000675.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_and_place_the_stapled_paper_into_a_new_stack/episode_8/cam_high/frame_000675.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_and_place_the_stapled_paper_into_a_new_stack/episode_8/cam_high/frame_000675.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_and_place_the_stapled_paper_into_a_new_stack/episode_8/cam_high/frame_000650.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_and_place_the_stapled_paper_into_a_new_stack/episode_8/cam_high/frame_000650.jpg", "images/human_egodex/Pick_up_one_piece_of_paper_from_a_stack_staple_it_with_a_stapler_and_place_the_stapled_paper_into_a_new_stack/episode_8/cam_high/frame_000650.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up one piece of paper from a stack staple it with a stapler and place the stapled paper into a new stack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up one piece of paper from a stack staple it with a stapler and place the stapled paper into a new stack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.7, "hop_absolute_value": 3.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Insert_the_black_USB_into_the_USB_bank_episode_0-train-bf_000100-af_000075-minus-0000018", "task": "Insert the black USB into the USB bank", "image": [ "images/human_egodex/Insert_the_black_USB_into_the_USB_bank/episode_0/cam_high/frame_000000.jpg", "images/human_egodex/Insert_the_black_USB_into_the_USB_bank/episode_0/cam_high/frame_000162.jpg", "images/human_egodex/Insert_the_black_USB_into_the_USB_bank/episode_0/cam_high/frame_000100.jpg", "images/human_egodex/Insert_the_black_USB_into_the_USB_bank/episode_0/cam_high/frame_000100.jpg", "images/human_egodex/Insert_the_black_USB_into_the_USB_bank/episode_0/cam_high/frame_000100.jpg", "images/human_egodex/Insert_the_black_USB_into_the_USB_bank/episode_0/cam_high/frame_000075.jpg", "images/human_egodex/Insert_the_black_USB_into_the_USB_bank/episode_0/cam_high/frame_000075.jpg", "images/human_egodex/Insert_the_black_USB_into_the_USB_bank/episode_0/cam_high/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert the black USB into the USB bank`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert the black USB into the USB bank`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Crumple_a_piece_of_paper_on_a_pink_tablecloth_with_a_lavender_background_episode_13-train-bf_000200-af_000175-minus-0000052", "task": "Crumple a piece of paper on a pink tablecloth with a lavender background", "image": [ "images/human_egodex/Crumple_a_piece_of_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_4/cam_high/frame_000000.jpg", "images/human_egodex/Crumple_a_piece_of_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_4/cam_high/frame_000321.jpg", "images/human_egodex/Crumple_a_piece_of_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_13/cam_high/frame_000200.jpg", "images/human_egodex/Crumple_a_piece_of_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_13/cam_high/frame_000200.jpg", "images/human_egodex/Crumple_a_piece_of_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_13/cam_high/frame_000200.jpg", "images/human_egodex/Crumple_a_piece_of_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_13/cam_high/frame_000175.jpg", "images/human_egodex/Crumple_a_piece_of_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_13/cam_high/frame_000175.jpg", "images/human_egodex/Crumple_a_piece_of_paper_on_a_pink_tablecloth_with_a_lavender_background/episode_13/cam_high/frame_000175.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Crumple a piece of paper on a pink tablecloth with a lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Crumple a piece of paper on a pink tablecloth with a lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Make_a_sandwich_using_breads_lettuce_cucumbers_tomato_cheese_turkey_and_onion_while_sitting_at_a_table_with_a_blue_background_episode_6-train-bf_000300-af_000275-minus-0000123", "task": "Make a sandwich using breads lettuce cucumbers tomato cheese turkey and onion while sitting at a table with a blue background", "image": [ "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_tomato_cheese_turkey_and_onion_while_sitting_at_a_table_with_a_blue_background/episode_1/cam_high/frame_000000.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_tomato_cheese_turkey_and_onion_while_sitting_at_a_table_with_a_blue_background/episode_1/cam_high/frame_000438.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_tomato_cheese_turkey_and_onion_while_sitting_at_a_table_with_a_blue_background/episode_6/cam_high/frame_000300.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_tomato_cheese_turkey_and_onion_while_sitting_at_a_table_with_a_blue_background/episode_6/cam_high/frame_000300.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_tomato_cheese_turkey_and_onion_while_sitting_at_a_table_with_a_blue_background/episode_6/cam_high/frame_000300.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_tomato_cheese_turkey_and_onion_while_sitting_at_a_table_with_a_blue_background/episode_6/cam_high/frame_000275.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_tomato_cheese_turkey_and_onion_while_sitting_at_a_table_with_a_blue_background/episode_6/cam_high/frame_000275.jpg", "images/human_egodex/Make_a_sandwich_using_breads_lettuce_cucumbers_tomato_cheese_turkey_and_onion_while_sitting_at_a_table_with_a_blue_background/episode_6/cam_high/frame_000275.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Make a sandwich using breads lettuce cucumbers tomato cheese turkey and onion while sitting at a table with a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Make a sandwich using breads lettuce cucumbers tomato cheese turkey and onion while sitting at a table with a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -8.3, "hop_absolute_value": 8.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Remove_the_polar_bear_or_accessories_from_the_crystal_slime_reshape_it_and_return_it_to_its_original_container_episode_4-train-bf_000250-af_000050-minus-0000018", "task": "Remove the polar bear or accessories from the crystal slime reshape it and return it to its original container", "image": [ "images/human_egodex/Remove_the_polar_bear_or_accessories_from_the_crystal_slime_reshape_it_and_return_it_to_its_original_container/episode_4/cam_high/frame_000000.jpg", "images/human_egodex/Remove_the_polar_bear_or_accessories_from_the_crystal_slime_reshape_it_and_return_it_to_its_original_container/episode_4/cam_high/frame_000299.jpg", "images/human_egodex/Remove_the_polar_bear_or_accessories_from_the_crystal_slime_reshape_it_and_return_it_to_its_original_container/episode_4/cam_high/frame_000250.jpg", "images/human_egodex/Remove_the_polar_bear_or_accessories_from_the_crystal_slime_reshape_it_and_return_it_to_its_original_container/episode_4/cam_high/frame_000250.jpg", "images/human_egodex/Remove_the_polar_bear_or_accessories_from_the_crystal_slime_reshape_it_and_return_it_to_its_original_container/episode_4/cam_high/frame_000250.jpg", "images/human_egodex/Remove_the_polar_bear_or_accessories_from_the_crystal_slime_reshape_it_and_return_it_to_its_original_container/episode_4/cam_high/frame_000050.jpg", "images/human_egodex/Remove_the_polar_bear_or_accessories_from_the_crystal_slime_reshape_it_and_return_it_to_its_original_container/episode_4/cam_high/frame_000050.jpg", "images/human_egodex/Remove_the_polar_bear_or_accessories_from_the_crystal_slime_reshape_it_and_return_it_to_its_original_container/episode_4/cam_high/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Remove the polar bear or accessories from the crystal slime reshape it and return it to its original container`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Remove the polar bear or accessories from the crystal slime reshape it and return it to its original container`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Arrange_dominoes_upright_on_a_metal_table_and_topple_them_episode_9-train-bf_001875-af_001850-minus-0000326", "task": "Arrange dominoes upright on a metal table and topple them", "image": [ "images/human_egodex/Arrange_dominoes_upright_on_a_metal_table_and_topple_them/episode_9/cam_high/frame_000000.jpg", "images/human_egodex/Arrange_dominoes_upright_on_a_metal_table_and_topple_them/episode_9/cam_high/frame_001962.jpg", "images/human_egodex/Arrange_dominoes_upright_on_a_metal_table_and_topple_them/episode_9/cam_high/frame_001875.jpg", "images/human_egodex/Arrange_dominoes_upright_on_a_metal_table_and_topple_them/episode_9/cam_high/frame_001875.jpg", "images/human_egodex/Arrange_dominoes_upright_on_a_metal_table_and_topple_them/episode_9/cam_high/frame_001875.jpg", "images/human_egodex/Arrange_dominoes_upright_on_a_metal_table_and_topple_them/episode_9/cam_high/frame_001850.jpg", "images/human_egodex/Arrange_dominoes_upright_on_a_metal_table_and_topple_them/episode_9/cam_high/frame_001850.jpg", "images/human_egodex/Arrange_dominoes_upright_on_a_metal_table_and_topple_them/episode_9/cam_high/frame_001850.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Arrange dominoes upright on a metal table and topple them`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Arrange dominoes upright on a metal table and topple them`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.3, "hop_absolute_value": 1.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Disassemble_a_chair_while_sitting_against_a_lavender_background_episode_16-train-bf_000075-af_000050-minus-0000017", "task": "Disassemble a chair while sitting against a lavender background", "image": [ "images/human_egodex/Disassemble_a_chair_while_sitting_against_a_lavender_background/episode_13/cam_high/frame_000000.jpg", "images/human_egodex/Disassemble_a_chair_while_sitting_against_a_lavender_background/episode_13/cam_high/frame_000071.jpg", "images/human_egodex/Disassemble_a_chair_while_sitting_against_a_lavender_background/episode_16/cam_high/frame_000075.jpg", "images/human_egodex/Disassemble_a_chair_while_sitting_against_a_lavender_background/episode_16/cam_high/frame_000075.jpg", "images/human_egodex/Disassemble_a_chair_while_sitting_against_a_lavender_background/episode_16/cam_high/frame_000075.jpg", "images/human_egodex/Disassemble_a_chair_while_sitting_against_a_lavender_background/episode_16/cam_high/frame_000050.jpg", "images/human_egodex/Disassemble_a_chair_while_sitting_against_a_lavender_background/episode_16/cam_high/frame_000050.jpg", "images/human_egodex/Disassemble_a_chair_while_sitting_against_a_lavender_background/episode_16/cam_high/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Disassemble a chair while sitting against a lavender background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Disassemble a chair while sitting against a lavender background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Collect_the_rolled_dice_back_into_one_hand_from_the_brown_background_episode_11-train-bf_000100-af_000000-minus-0000003", "task": "Collect the rolled dice back into one hand from the brown background", "image": [ "images/human_egodex/Collect_the_rolled_dice_back_into_one_hand_from_the_brown_background/episode_6/cam_high/frame_000000.jpg", "images/human_egodex/Collect_the_rolled_dice_back_into_one_hand_from_the_brown_background/episode_6/cam_high/frame_000215.jpg", "images/human_egodex/Collect_the_rolled_dice_back_into_one_hand_from_the_brown_background/episode_11/cam_high/frame_000100.jpg", "images/human_egodex/Collect_the_rolled_dice_back_into_one_hand_from_the_brown_background/episode_11/cam_high/frame_000100.jpg", "images/human_egodex/Collect_the_rolled_dice_back_into_one_hand_from_the_brown_background/episode_11/cam_high/frame_000100.jpg", "images/human_egodex/Collect_the_rolled_dice_back_into_one_hand_from_the_brown_background/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Collect_the_rolled_dice_back_into_one_hand_from_the_brown_background/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Collect_the_rolled_dice_back_into_one_hand_from_the_brown_background/episode_11/cam_high/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Collect the rolled dice back into one hand from the brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Collect the rolled dice back into one hand from the brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Distribute_pebbles_from_the_bank_into_the_board_while_sitting_at_a_wooden_tablecloth_with_a_white_background_episode_11-train-bf_000725-af_000000-minus-0000018", "task": "Distribute pebbles from the bank into the board while sitting at a wooden tablecloth with a white background", "image": [ "images/human_egodex/Distribute_pebbles_from_the_bank_into_the_board_while_sitting_at_a_wooden_tablecloth_with_a_white_background/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Distribute_pebbles_from_the_bank_into_the_board_while_sitting_at_a_wooden_tablecloth_with_a_white_background/episode_7/cam_high/frame_000581.jpg", "images/human_egodex/Distribute_pebbles_from_the_bank_into_the_board_while_sitting_at_a_wooden_tablecloth_with_a_white_background/episode_11/cam_high/frame_000725.jpg", "images/human_egodex/Distribute_pebbles_from_the_bank_into_the_board_while_sitting_at_a_wooden_tablecloth_with_a_white_background/episode_11/cam_high/frame_000725.jpg", "images/human_egodex/Distribute_pebbles_from_the_bank_into_the_board_while_sitting_at_a_wooden_tablecloth_with_a_white_background/episode_11/cam_high/frame_000725.jpg", "images/human_egodex/Distribute_pebbles_from_the_bank_into_the_board_while_sitting_at_a_wooden_tablecloth_with_a_white_background/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Distribute_pebbles_from_the_bank_into_the_board_while_sitting_at_a_wooden_tablecloth_with_a_white_background/episode_11/cam_high/frame_000000.jpg", "images/human_egodex/Distribute_pebbles_from_the_bank_into_the_board_while_sitting_at_a_wooden_tablecloth_with_a_white_background/episode_11/cam_high/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Distribute pebbles from the bank into the board while sitting at a wooden tablecloth with a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Distribute pebbles from the bank into the board while sitting at a wooden tablecloth with a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Insert_a_USB_into_the_right_hand_while_sitting_with_a_blue_background_episode_15-train-bf_000425-af_000400-minus-0000272", "task": "Insert a USB into the right hand while sitting with a blue background", "image": [ "images/human_egodex/Insert_a_USB_into_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000000.jpg", "images/human_egodex/Insert_a_USB_into_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000545.jpg", "images/human_egodex/Insert_a_USB_into_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000425.jpg", "images/human_egodex/Insert_a_USB_into_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000425.jpg", "images/human_egodex/Insert_a_USB_into_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000425.jpg", "images/human_egodex/Insert_a_USB_into_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000400.jpg", "images/human_egodex/Insert_a_USB_into_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000400.jpg", "images/human_egodex/Insert_a_USB_into_the_right_hand_while_sitting_with_a_blue_background/episode_15/cam_high/frame_000400.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Insert a USB into the right hand while sitting with a blue background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Insert a USB into the right hand while sitting with a blue background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.9, "hop_absolute_value": 5.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Braid_the_three_strands_of_salmonpink_yarn_braid_while_sitting_with_a_brown_background_episode_2-train-bf_000575-af_000550-minus-0000375", "task": "Braid the three strands of salmonpink yarn braid while sitting with a brown background", "image": [ "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn_braid_while_sitting_with_a_brown_background/episode_1/cam_high/frame_000000.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn_braid_while_sitting_with_a_brown_background/episode_1/cam_high/frame_000966.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn_braid_while_sitting_with_a_brown_background/episode_2/cam_high/frame_000575.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn_braid_while_sitting_with_a_brown_background/episode_2/cam_high/frame_000575.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn_braid_while_sitting_with_a_brown_background/episode_2/cam_high/frame_000575.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn_braid_while_sitting_with_a_brown_background/episode_2/cam_high/frame_000550.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn_braid_while_sitting_with_a_brown_background/episode_2/cam_high/frame_000550.jpg", "images/human_egodex/Braid_the_three_strands_of_salmonpink_yarn_braid_while_sitting_with_a_brown_background/episode_2/cam_high/frame_000550.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Braid the three strands of salmonpink yarn braid while sitting with a brown background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Braid the three strands of salmonpink yarn braid while sitting with a brown background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.3, "hop_absolute_value": 4.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Take_toothbrush_from_gray_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_gray_cup_episode_7-train-bf_000125-af_000100-minus-0000135", "task": "Take toothbrush from gray cup unscrew toothpaste cap put toothpaste on toothbrush screw cap back on toothpaste put toothpaste down and put toothbrush back into gray cup", "image": [ "images/human_egodex/Take_toothbrush_from_gray_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_gray_cup/episode_7/cam_high/frame_000000.jpg", "images/human_egodex/Take_toothbrush_from_gray_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_gray_cup/episode_7/cam_high/frame_000654.jpg", "images/human_egodex/Take_toothbrush_from_gray_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_gray_cup/episode_7/cam_high/frame_000125.jpg", "images/human_egodex/Take_toothbrush_from_gray_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_gray_cup/episode_7/cam_high/frame_000125.jpg", "images/human_egodex/Take_toothbrush_from_gray_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_gray_cup/episode_7/cam_high/frame_000125.jpg", "images/human_egodex/Take_toothbrush_from_gray_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_gray_cup/episode_7/cam_high/frame_000100.jpg", "images/human_egodex/Take_toothbrush_from_gray_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_gray_cup/episode_7/cam_high/frame_000100.jpg", "images/human_egodex/Take_toothbrush_from_gray_cup_unscrew_toothpaste_cap_put_toothpaste_on_toothbrush_screw_cap_back_on_toothpaste_put_toothpaste_down_and_put_toothbrush_back_into_gray_cup/episode_7/cam_high/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take toothbrush from gray cup unscrew toothpaste cap put toothpaste on toothbrush screw cap back on toothpaste put toothpaste down and put toothbrush back into gray cup`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take toothbrush from gray cup unscrew toothpaste cap put toothpaste on toothbrush screw cap back on toothpaste put toothpaste down and put toothbrush back into gray cup`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Set_up_chess_pieces_on_a_chessboard_placed_on_a_wooden_table_while_sitting_against_a_white_background_episode_8-train-bf_001850-af_001825-minus-0000481", "task": "Set up chess pieces on a chessboard placed on a wooden table while sitting against a white background", "image": [ "images/human_egodex/Set_up_chess_pieces_on_a_chessboard_placed_on_a_wooden_table_while_sitting_against_a_white_background/episode_8/cam_high/frame_000000.jpg", "images/human_egodex/Set_up_chess_pieces_on_a_chessboard_placed_on_a_wooden_table_while_sitting_against_a_white_background/episode_8/cam_high/frame_002465.jpg", "images/human_egodex/Set_up_chess_pieces_on_a_chessboard_placed_on_a_wooden_table_while_sitting_against_a_white_background/episode_8/cam_high/frame_001850.jpg", "images/human_egodex/Set_up_chess_pieces_on_a_chessboard_placed_on_a_wooden_table_while_sitting_against_a_white_background/episode_8/cam_high/frame_001850.jpg", "images/human_egodex/Set_up_chess_pieces_on_a_chessboard_placed_on_a_wooden_table_while_sitting_against_a_white_background/episode_8/cam_high/frame_001850.jpg", "images/human_egodex/Set_up_chess_pieces_on_a_chessboard_placed_on_a_wooden_table_while_sitting_against_a_white_background/episode_8/cam_high/frame_001825.jpg", "images/human_egodex/Set_up_chess_pieces_on_a_chessboard_placed_on_a_wooden_table_while_sitting_against_a_white_background/episode_8/cam_high/frame_001825.jpg", "images/human_egodex/Set_up_chess_pieces_on_a_chessboard_placed_on_a_wooden_table_while_sitting_against_a_white_background/episode_8/cam_high/frame_001825.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Set up chess pieces on a chessboard placed on a wooden table while sitting against a white background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Set up chess pieces on a chessboard placed on a wooden table while sitting against a white background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -1.4, "hop_absolute_value": 1.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_egodex" }, { "id": "human_egodex_Assemble_jenga_blocks_into_a_tower_on_a_blue_tablecloth_with_a_pink_background_episode_5-train-bf_000175-af_000000-minus-0000005", "task": "Assemble jenga blocks into a tower on a blue tablecloth with a pink background", "image": [ "images/human_egodex/Assemble_jenga_blocks_into_a_tower_on_a_blue_tablecloth_with_a_pink_background/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Assemble_jenga_blocks_into_a_tower_on_a_blue_tablecloth_with_a_pink_background/episode_5/cam_high/frame_000299.jpg", "images/human_egodex/Assemble_jenga_blocks_into_a_tower_on_a_blue_tablecloth_with_a_pink_background/episode_5/cam_high/frame_000175.jpg", "images/human_egodex/Assemble_jenga_blocks_into_a_tower_on_a_blue_tablecloth_with_a_pink_background/episode_5/cam_high/frame_000175.jpg", "images/human_egodex/Assemble_jenga_blocks_into_a_tower_on_a_blue_tablecloth_with_a_pink_background/episode_5/cam_high/frame_000175.jpg", "images/human_egodex/Assemble_jenga_blocks_into_a_tower_on_a_blue_tablecloth_with_a_pink_background/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Assemble_jenga_blocks_into_a_tower_on_a_blue_tablecloth_with_a_pink_background/episode_5/cam_high/frame_000000.jpg", "images/human_egodex/Assemble_jenga_blocks_into_a_tower_on_a_blue_tablecloth_with_a_pink_background/episode_5/cam_high/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Assemble jenga blocks into a tower on a blue tablecloth with a pink background`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Assemble jenga blocks into a tower on a blue tablecloth with a pink background`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_egodex" }, { "id": "human_pika_pour_bowl_episode_21-train-bf_000315-af_000360-plus-0000873", "task": "pour bowl", "image": [ "images/human_pika/pour_bowl/episode_3/cam_high/frame_000000.jpg", "images/human_pika/pour_bowl/episode_3/cam_high/frame_000895.jpg", "images/human_pika/pour_bowl/episode_21/cam_high/frame_000315.jpg", "images/human_pika/pour_bowl/episode_21/cam_left_wrist/frame_000315.jpg", "images/human_pika/pour_bowl/episode_21/cam_right_wrist/frame_000315.jpg", "images/human_pika/pour_bowl/episode_21/cam_high/frame_000360.jpg", "images/human_pika/pour_bowl/episode_21/cam_left_wrist/frame_000360.jpg", "images/human_pika/pour_bowl/episode_21/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 7.3, "hop_absolute_value": 7.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_place_test_tube_episode_15-train-bf_000015-af_000120-plus-0000677", "task": "place test tube", "image": [ "images/human_pika/place_test_tube/episode_15/cam_high/frame_000000.jpg", "images/human_pika/place_test_tube/episode_15/cam_high/frame_000454.jpg", "images/human_pika/place_test_tube/episode_15/cam_high/frame_000015.jpg", "images/human_pika/place_test_tube/episode_15/cam_left_wrist/frame_000015.jpg", "images/human_pika/place_test_tube/episode_15/cam_right_wrist/frame_000015.jpg", "images/human_pika/place_test_tube/episode_15/cam_high/frame_000120.jpg", "images/human_pika/place_test_tube/episode_15/cam_left_wrist/frame_000120.jpg", "images/human_pika/place_test_tube/episode_15/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`place test tube`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `place test tube`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 24.1, "hop_absolute_value": 24.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_groceries_into_bag_episode_5-train-bf_000015-af_000255-plus-0000934", "task": "put groceries into bag", "image": [ "images/human_pika/put_groceries_into_bag/episode_13/cam_high/frame_000000.jpg", "images/human_pika/put_groceries_into_bag/episode_13/cam_high/frame_000455.jpg", "images/human_pika/put_groceries_into_bag/episode_5/cam_high/frame_000015.jpg", "images/human_pika/put_groceries_into_bag/episode_5/cam_left_wrist/frame_000015.jpg", "images/human_pika/put_groceries_into_bag/episode_5/cam_right_wrist/frame_000015.jpg", "images/human_pika/put_groceries_into_bag/episode_5/cam_high/frame_000255.jpg", "images/human_pika/put_groceries_into_bag/episode_5/cam_left_wrist/frame_000255.jpg", "images/human_pika/put_groceries_into_bag/episode_5/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put groceries into bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put groceries into bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 48.5, "hop_absolute_value": 48.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bowl_episode_17-train-bf_000345-af_000570-plus-0000388", "task": "pour bowl", "image": [ "images/human_pika/pour_bowl/episode_10/cam_high/frame_000000.jpg", "images/human_pika/pour_bowl/episode_10/cam_high/frame_000893.jpg", "images/human_pika/pour_bowl/episode_17/cam_high/frame_000345.jpg", "images/human_pika/pour_bowl/episode_17/cam_left_wrist/frame_000345.jpg", "images/human_pika/pour_bowl/episode_17/cam_right_wrist/frame_000345.jpg", "images/human_pika/pour_bowl/episode_17/cam_high/frame_000570.jpg", "images/human_pika/pour_bowl/episode_17/cam_left_wrist/frame_000570.jpg", "images/human_pika/pour_bowl/episode_17/cam_right_wrist/frame_000570.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 46.9, "hop_absolute_value": 46.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_groceries_into_bag_episode_6-train-bf_000180-af_000225-plus-0000293", "task": "put groceries into bag", "image": [ "images/human_pika/put_groceries_into_bag/episode_9/cam_high/frame_000000.jpg", "images/human_pika/put_groceries_into_bag/episode_9/cam_high/frame_000463.jpg", "images/human_pika/put_groceries_into_bag/episode_6/cam_high/frame_000180.jpg", "images/human_pika/put_groceries_into_bag/episode_6/cam_left_wrist/frame_000180.jpg", "images/human_pika/put_groceries_into_bag/episode_6/cam_right_wrist/frame_000180.jpg", "images/human_pika/put_groceries_into_bag/episode_6/cam_high/frame_000225.jpg", "images/human_pika/put_groceries_into_bag/episode_6/cam_left_wrist/frame_000225.jpg", "images/human_pika/put_groceries_into_bag/episode_6/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put groceries into bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put groceries into bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_groceries_into_bag_episode_3-train-bf_000015-af_000120-plus-0000771", "task": "put groceries into bag", "image": [ "images/human_pika/put_groceries_into_bag/episode_12/cam_high/frame_000000.jpg", "images/human_pika/put_groceries_into_bag/episode_12/cam_high/frame_000492.jpg", "images/human_pika/put_groceries_into_bag/episode_3/cam_high/frame_000015.jpg", "images/human_pika/put_groceries_into_bag/episode_3/cam_left_wrist/frame_000015.jpg", "images/human_pika/put_groceries_into_bag/episode_3/cam_right_wrist/frame_000015.jpg", "images/human_pika/put_groceries_into_bag/episode_3/cam_high/frame_000120.jpg", "images/human_pika/put_groceries_into_bag/episode_3/cam_left_wrist/frame_000120.jpg", "images/human_pika/put_groceries_into_bag/episode_3/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put groceries into bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put groceries into bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.6, "hop_absolute_value": 20.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bowl_episode_17-train-bf_000345-af_000705-plus-0000475", "task": "pour bowl", "image": [ "images/human_pika/pour_bowl/episode_17/cam_high/frame_000000.jpg", "images/human_pika/pour_bowl/episode_17/cam_high/frame_000834.jpg", "images/human_pika/pour_bowl/episode_17/cam_high/frame_000345.jpg", "images/human_pika/pour_bowl/episode_17/cam_left_wrist/frame_000345.jpg", "images/human_pika/pour_bowl/episode_17/cam_right_wrist/frame_000345.jpg", "images/human_pika/pour_bowl/episode_17/cam_high/frame_000705.jpg", "images/human_pika/pour_bowl/episode_17/cam_left_wrist/frame_000705.jpg", "images/human_pika/pour_bowl/episode_17/cam_right_wrist/frame_000705.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_peach_to_the_basket_episode_23-train-bf_000045-af_000090-plus-0000572", "task": "put peach to the basket", "image": [ "images/human_pika/put_peach_to_the_basket/episode_1/cam_high/frame_000000.jpg", "images/human_pika/put_peach_to_the_basket/episode_1/cam_high/frame_000547.jpg", "images/human_pika/put_peach_to_the_basket/episode_23/cam_high/frame_000045.jpg", "images/human_pika/put_peach_to_the_basket/episode_23/cam_left_wrist/frame_000045.jpg", "images/human_pika/put_peach_to_the_basket/episode_23/cam_right_wrist/frame_000045.jpg", "images/human_pika/put_peach_to_the_basket/episode_23/cam_high/frame_000090.jpg", "images/human_pika/put_peach_to_the_basket/episode_23/cam_left_wrist/frame_000090.jpg", "images/human_pika/put_peach_to_the_basket/episode_23/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put peach to the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put peach to the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_cup_water_containing_episode_3-train-bf_000090-af_000240-plus-0000334", "task": "pour cup water containing", "image": [ "images/human_pika/pour_cup_water_containing/episode_21/cam_high/frame_000000.jpg", "images/human_pika/pour_cup_water_containing/episode_21/cam_high/frame_000566.jpg", "images/human_pika/pour_cup_water_containing/episode_3/cam_high/frame_000090.jpg", "images/human_pika/pour_cup_water_containing/episode_3/cam_left_wrist/frame_000090.jpg", "images/human_pika/pour_cup_water_containing/episode_3/cam_right_wrist/frame_000090.jpg", "images/human_pika/pour_cup_water_containing/episode_3/cam_high/frame_000240.jpg", "images/human_pika/pour_cup_water_containing/episode_3/cam_left_wrist/frame_000240.jpg", "images/human_pika/pour_cup_water_containing/episode_3/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour cup water containing`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour cup water containing`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 34.5, "hop_absolute_value": 34.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_place_test_tube_episode_7-train-bf_000030-af_000105-plus-0000640", "task": "place test tube", "image": [ "images/human_pika/place_test_tube/episode_14/cam_high/frame_000000.jpg", "images/human_pika/place_test_tube/episode_14/cam_high/frame_000508.jpg", "images/human_pika/place_test_tube/episode_7/cam_high/frame_000030.jpg", "images/human_pika/place_test_tube/episode_7/cam_left_wrist/frame_000030.jpg", "images/human_pika/place_test_tube/episode_7/cam_right_wrist/frame_000030.jpg", "images/human_pika/place_test_tube/episode_7/cam_high/frame_000105.jpg", "images/human_pika/place_test_tube/episode_7/cam_left_wrist/frame_000105.jpg", "images/human_pika/place_test_tube/episode_7/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`place test tube`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `place test tube`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 17.2, "hop_absolute_value": 17.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_stir_coffee_episode_10-train-bf_000180-af_000525-plus-0000563", "task": "stir coffee", "image": [ "images/human_pika/stir_coffee/episode_14/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_14/cam_high/frame_000652.jpg", "images/human_pika/stir_coffee/episode_10/cam_high/frame_000180.jpg", "images/human_pika/stir_coffee/episode_10/cam_left_wrist/frame_000180.jpg", "images/human_pika/stir_coffee/episode_10/cam_right_wrist/frame_000180.jpg", "images/human_pika/stir_coffee/episode_10/cam_high/frame_000525.jpg", "images/human_pika/stir_coffee/episode_10/cam_left_wrist/frame_000525.jpg", "images/human_pika/stir_coffee/episode_10/cam_right_wrist/frame_000525.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stir coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stir coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 92.0, "hop_absolute_value": 92.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_wipe_stains_episode_22-train-bf_000195-af_000345-plus-0000928", "task": "wipe stains", "image": [ "images/human_pika/wipe_stains/episode_21/cam_high/frame_000000.jpg", "images/human_pika/wipe_stains/episode_21/cam_high/frame_000524.jpg", "images/human_pika/wipe_stains/episode_22/cam_high/frame_000195.jpg", "images/human_pika/wipe_stains/episode_22/cam_left_wrist/frame_000195.jpg", "images/human_pika/wipe_stains/episode_22/cam_right_wrist/frame_000195.jpg", "images/human_pika/wipe_stains/episode_22/cam_high/frame_000345.jpg", "images/human_pika/wipe_stains/episode_22/cam_left_wrist/frame_000345.jpg", "images/human_pika/wipe_stains/episode_22/cam_right_wrist/frame_000345.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`wipe stains`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `wipe stains`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 43.5, "hop_absolute_value": 43.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_groceries_into_bag_episode_19-train-bf_000030-af_000315-plus-0000404", "task": "put groceries into bag", "image": [ "images/human_pika/put_groceries_into_bag/episode_23/cam_high/frame_000000.jpg", "images/human_pika/put_groceries_into_bag/episode_23/cam_high/frame_000450.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_high/frame_000030.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_left_wrist/frame_000030.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_right_wrist/frame_000030.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_high/frame_000315.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_left_wrist/frame_000315.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_right_wrist/frame_000315.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put groceries into bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put groceries into bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.6, "hop_absolute_value": 57.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bowl_episode_10-train-bf_000210-af_000240-plus-0000911", "task": "pour bowl", "image": [ "images/human_pika/pour_bowl/episode_10/cam_high/frame_000000.jpg", "images/human_pika/pour_bowl/episode_10/cam_high/frame_000893.jpg", "images/human_pika/pour_bowl/episode_10/cam_high/frame_000210.jpg", "images/human_pika/pour_bowl/episode_10/cam_left_wrist/frame_000210.jpg", "images/human_pika/pour_bowl/episode_10/cam_right_wrist/frame_000210.jpg", "images/human_pika/pour_bowl/episode_10/cam_high/frame_000240.jpg", "images/human_pika/pour_bowl/episode_10/cam_left_wrist/frame_000240.jpg", "images/human_pika/pour_bowl/episode_10/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.4, "hop_absolute_value": 4.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_test_tube_to_the_shelf_episode_17-train-bf_000120-af_000405-plus-0000401", "task": "put test tube to the shelf", "image": [ "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_high/frame_000000.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_high/frame_000621.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_high/frame_000120.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_left_wrist/frame_000120.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_right_wrist/frame_000120.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_high/frame_000405.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_left_wrist/frame_000405.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put test tube to the shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put test tube to the shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.6, "hop_absolute_value": 57.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_20-train-bf_000030-af_000165-plus-0000820", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_20/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_20/cam_high/frame_000552.jpg", "images/human_pika/pour_tea/episode_20/cam_high/frame_000030.jpg", "images/human_pika/pour_tea/episode_20/cam_left_wrist/frame_000030.jpg", "images/human_pika/pour_tea/episode_20/cam_right_wrist/frame_000030.jpg", "images/human_pika/pour_tea/episode_20/cam_high/frame_000165.jpg", "images/human_pika/pour_tea/episode_20/cam_left_wrist/frame_000165.jpg", "images/human_pika/pour_tea/episode_20/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 26.5, "hop_absolute_value": 26.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_banana_episode_12-train-bf_000390-af_000525-plus-0000449", "task": "pour basket with banana", "image": [ "images/human_pika/pour_basket_with_banana/episode_11/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_banana/episode_11/cam_high/frame_000552.jpg", "images/human_pika/pour_basket_with_banana/episode_12/cam_high/frame_000390.jpg", "images/human_pika/pour_basket_with_banana/episode_12/cam_left_wrist/frame_000390.jpg", "images/human_pika/pour_basket_with_banana/episode_12/cam_right_wrist/frame_000390.jpg", "images/human_pika/pour_basket_with_banana/episode_12/cam_high/frame_000525.jpg", "images/human_pika/pour_basket_with_banana/episode_12/cam_left_wrist/frame_000525.jpg", "images/human_pika/pour_basket_with_banana/episode_12/cam_right_wrist/frame_000525.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with banana`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with banana`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 69.2, "hop_absolute_value": 69.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_scoop_bean_episode_5-train-bf_000090-af_000510-plus-0000535", "task": "scoop bean", "image": [ "images/human_pika/scoop_bean/episode_8/cam_high/frame_000000.jpg", "images/human_pika/scoop_bean/episode_8/cam_high/frame_000486.jpg", "images/human_pika/scoop_bean/episode_5/cam_high/frame_000090.jpg", "images/human_pika/scoop_bean/episode_5/cam_left_wrist/frame_000090.jpg", "images/human_pika/scoop_bean/episode_5/cam_right_wrist/frame_000090.jpg", "images/human_pika/scoop_bean/episode_5/cam_high/frame_000510.jpg", "images/human_pika/scoop_bean/episode_5/cam_left_wrist/frame_000510.jpg", "images/human_pika/scoop_bean/episode_5/cam_right_wrist/frame_000510.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`scoop bean`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `scoop bean`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 90.3, "hop_absolute_value": 90.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_towel_episode_23-train-bf_000120-af_000480-plus-0000539", "task": "fold towel", "image": [ "images/human_pika/fold_towel/episode_23/cam_high/frame_000000.jpg", "images/human_pika/fold_towel/episode_23/cam_high/frame_000505.jpg", "images/human_pika/fold_towel/episode_23/cam_high/frame_000120.jpg", "images/human_pika/fold_towel/episode_23/cam_left_wrist/frame_000120.jpg", "images/human_pika/fold_towel/episode_23/cam_right_wrist/frame_000120.jpg", "images/human_pika/fold_towel/episode_23/cam_high/frame_000480.jpg", "images/human_pika/fold_towel/episode_23/cam_left_wrist/frame_000480.jpg", "images/human_pika/fold_towel/episode_23/cam_right_wrist/frame_000480.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold towel`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold towel`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 96.0, "hop_absolute_value": 96.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_towel_episode_4-train-bf_000000-af_000255-plus-0000375", "task": "fold towel", "image": [ "images/human_pika/fold_towel/episode_7/cam_high/frame_000000.jpg", "images/human_pika/fold_towel/episode_7/cam_high/frame_000513.jpg", "images/human_pika/fold_towel/episode_4/cam_high/frame_000000.jpg", "images/human_pika/fold_towel/episode_4/cam_left_wrist/frame_000000.jpg", "images/human_pika/fold_towel/episode_4/cam_right_wrist/frame_000000.jpg", "images/human_pika/fold_towel/episode_4/cam_high/frame_000255.jpg", "images/human_pika/fold_towel/episode_4/cam_left_wrist/frame_000255.jpg", "images/human_pika/fold_towel/episode_4/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold towel`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold towel`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 51.5, "hop_absolute_value": 51.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_stir_coffee_episode_14-train-bf_000105-af_000120-plus-0000606", "task": "stir coffee", "image": [ "images/human_pika/stir_coffee/episode_22/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_22/cam_high/frame_000631.jpg", "images/human_pika/stir_coffee/episode_14/cam_high/frame_000105.jpg", "images/human_pika/stir_coffee/episode_14/cam_left_wrist/frame_000105.jpg", "images/human_pika/stir_coffee/episode_14/cam_right_wrist/frame_000105.jpg", "images/human_pika/stir_coffee/episode_14/cam_high/frame_000120.jpg", "images/human_pika/stir_coffee/episode_14/cam_left_wrist/frame_000120.jpg", "images/human_pika/stir_coffee/episode_14/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stir coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stir coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.8, "hop_absolute_value": 2.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_towel_episode_2-train-bf_000045-af_000300-plus-0000845", "task": "fold towel", "image": [ "images/human_pika/fold_towel/episode_1/cam_high/frame_000000.jpg", "images/human_pika/fold_towel/episode_1/cam_high/frame_000574.jpg", "images/human_pika/fold_towel/episode_2/cam_high/frame_000045.jpg", "images/human_pika/fold_towel/episode_2/cam_left_wrist/frame_000045.jpg", "images/human_pika/fold_towel/episode_2/cam_right_wrist/frame_000045.jpg", "images/human_pika/fold_towel/episode_2/cam_high/frame_000300.jpg", "images/human_pika/fold_towel/episode_2/cam_left_wrist/frame_000300.jpg", "images/human_pika/fold_towel/episode_2/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold towel`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold towel`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 63.0, "hop_absolute_value": 63.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_cup_water_containing_episode_4-train-bf_000420-af_000480-plus-0000460", "task": "pour cup water containing", "image": [ "images/human_pika/pour_cup_water_containing/episode_23/cam_high/frame_000000.jpg", "images/human_pika/pour_cup_water_containing/episode_23/cam_high/frame_000453.jpg", "images/human_pika/pour_cup_water_containing/episode_4/cam_high/frame_000420.jpg", "images/human_pika/pour_cup_water_containing/episode_4/cam_left_wrist/frame_000420.jpg", "images/human_pika/pour_cup_water_containing/episode_4/cam_right_wrist/frame_000420.jpg", "images/human_pika/pour_cup_water_containing/episode_4/cam_high/frame_000480.jpg", "images/human_pika/pour_cup_water_containing/episode_4/cam_left_wrist/frame_000480.jpg", "images/human_pika/pour_cup_water_containing/episode_4/cam_right_wrist/frame_000480.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour cup water containing`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour cup water containing`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_peach_episode_0-train-bf_000300-af_000525-plus-0000456", "task": "pour basket with peach", "image": [ "images/human_pika/pour_basket_with_peach/episode_14/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_peach/episode_14/cam_high/frame_000520.jpg", "images/human_pika/pour_basket_with_peach/episode_0/cam_high/frame_000300.jpg", "images/human_pika/pour_basket_with_peach/episode_0/cam_left_wrist/frame_000300.jpg", "images/human_pika/pour_basket_with_peach/episode_0/cam_right_wrist/frame_000300.jpg", "images/human_pika/pour_basket_with_peach/episode_0/cam_high/frame_000525.jpg", "images/human_pika/pour_basket_with_peach/episode_0/cam_left_wrist/frame_000525.jpg", "images/human_pika/pour_basket_with_peach/episode_0/cam_right_wrist/frame_000525.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_groceries_into_bag_episode_19-train-bf_000240-af_000330-plus-0000854", "task": "put groceries into bag", "image": [ "images/human_pika/put_groceries_into_bag/episode_16/cam_high/frame_000000.jpg", "images/human_pika/put_groceries_into_bag/episode_16/cam_high/frame_000515.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_high/frame_000240.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_left_wrist/frame_000240.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_right_wrist/frame_000240.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_high/frame_000330.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_left_wrist/frame_000330.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put groceries into bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put groceries into bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 31.6, "hop_absolute_value": 31.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_groceries_into_bag_episode_10-train-bf_000345-af_000450-plus-0000479", "task": "put groceries into bag", "image": [ "images/human_pika/put_groceries_into_bag/episode_10/cam_high/frame_000000.jpg", "images/human_pika/put_groceries_into_bag/episode_10/cam_high/frame_000467.jpg", "images/human_pika/put_groceries_into_bag/episode_10/cam_high/frame_000345.jpg", "images/human_pika/put_groceries_into_bag/episode_10/cam_left_wrist/frame_000345.jpg", "images/human_pika/put_groceries_into_bag/episode_10/cam_right_wrist/frame_000345.jpg", "images/human_pika/put_groceries_into_bag/episode_10/cam_high/frame_000450.jpg", "images/human_pika/put_groceries_into_bag/episode_10/cam_left_wrist/frame_000450.jpg", "images/human_pika/put_groceries_into_bag/episode_10/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put groceries into bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put groceries into bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_plate_storage_episode_7-train-bf_000165-af_000300-plus-0000345", "task": "pour plate storage", "image": [ "images/human_pika/pour_plate_storage/episode_13/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_13/cam_high/frame_000491.jpg", "images/human_pika/pour_plate_storage/episode_7/cam_high/frame_000165.jpg", "images/human_pika/pour_plate_storage/episode_7/cam_left_wrist/frame_000165.jpg", "images/human_pika/pour_plate_storage/episode_7/cam_right_wrist/frame_000165.jpg", "images/human_pika/pour_plate_storage/episode_7/cam_high/frame_000300.jpg", "images/human_pika/pour_plate_storage/episode_7/cam_left_wrist/frame_000300.jpg", "images/human_pika/pour_plate_storage/episode_7/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour plate storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour plate storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 37.5, "hop_absolute_value": 37.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_test_tube_to_the_shelf_episode_4-train-bf_000360-af_000405-plus-0000312", "task": "put test tube to the shelf", "image": [ "images/human_pika/put_test_tube_to_the_shelf/episode_9/cam_high/frame_000000.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_9/cam_high/frame_000508.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_4/cam_high/frame_000360.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_4/cam_left_wrist/frame_000360.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_4/cam_right_wrist/frame_000360.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_4/cam_high/frame_000405.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_4/cam_left_wrist/frame_000405.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_4/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put test tube to the shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put test tube to the shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 27.3, "hop_absolute_value": 27.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_cup_water_containing_episode_10-train-bf_000405-af_000468-plus-0000513", "task": "pour cup water containing", "image": [ "images/human_pika/pour_cup_water_containing/episode_22/cam_high/frame_000000.jpg", "images/human_pika/pour_cup_water_containing/episode_22/cam_high/frame_000563.jpg", "images/human_pika/pour_cup_water_containing/episode_10/cam_high/frame_000405.jpg", "images/human_pika/pour_cup_water_containing/episode_10/cam_left_wrist/frame_000405.jpg", "images/human_pika/pour_cup_water_containing/episode_10/cam_right_wrist/frame_000405.jpg", "images/human_pika/pour_cup_water_containing/episode_10/cam_high/frame_000468.jpg", "images/human_pika/pour_cup_water_containing/episode_10/cam_left_wrist/frame_000468.jpg", "images/human_pika/pour_cup_water_containing/episode_10/cam_right_wrist/frame_000468.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour cup water containing`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour cup water containing`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_orange_storage_episode_16-train-bf_000120-af_000165-plus-0000644", "task": "orange storage", "image": [ "images/human_pika/orange_storage/episode_18/cam_high/frame_000000.jpg", "images/human_pika/orange_storage/episode_18/cam_high/frame_000518.jpg", "images/human_pika/orange_storage/episode_16/cam_high/frame_000120.jpg", "images/human_pika/orange_storage/episode_16/cam_left_wrist/frame_000120.jpg", "images/human_pika/orange_storage/episode_16/cam_right_wrist/frame_000120.jpg", "images/human_pika/orange_storage/episode_16/cam_high/frame_000165.jpg", "images/human_pika/orange_storage/episode_16/cam_left_wrist/frame_000165.jpg", "images/human_pika/orange_storage/episode_16/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`orange storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `orange storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.3, "hop_absolute_value": 10.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_orange_storage_episode_11-train-bf_000015-af_000030-plus-0000289", "task": "orange storage", "image": [ "images/human_pika/orange_storage/episode_0/cam_high/frame_000000.jpg", "images/human_pika/orange_storage/episode_0/cam_high/frame_000694.jpg", "images/human_pika/orange_storage/episode_11/cam_high/frame_000015.jpg", "images/human_pika/orange_storage/episode_11/cam_left_wrist/frame_000015.jpg", "images/human_pika/orange_storage/episode_11/cam_right_wrist/frame_000015.jpg", "images/human_pika/orange_storage/episode_11/cam_high/frame_000030.jpg", "images/human_pika/orange_storage/episode_11/cam_left_wrist/frame_000030.jpg", "images/human_pika/orange_storage/episode_11/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`orange storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `orange storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.7, "hop_absolute_value": 2.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_wipe_stains_episode_24-train-bf_000030-af_000285-plus-0000383", "task": "wipe stains", "image": [ "images/human_pika/wipe_stains/episode_22/cam_high/frame_000000.jpg", "images/human_pika/wipe_stains/episode_22/cam_high/frame_000551.jpg", "images/human_pika/wipe_stains/episode_24/cam_high/frame_000030.jpg", "images/human_pika/wipe_stains/episode_24/cam_left_wrist/frame_000030.jpg", "images/human_pika/wipe_stains/episode_24/cam_right_wrist/frame_000030.jpg", "images/human_pika/wipe_stains/episode_24/cam_high/frame_000285.jpg", "images/human_pika/wipe_stains/episode_24/cam_left_wrist/frame_000285.jpg", "images/human_pika/wipe_stains/episode_24/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`wipe stains`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `wipe stains`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 44.7, "hop_absolute_value": 44.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_build_the_blocks_episode_2-train-bf_000240-af_000285-plus-0000711", "task": "build the blocks", "image": [ "images/human_pika/build_the_blocks/episode_16/cam_high/frame_000000.jpg", "images/human_pika/build_the_blocks/episode_16/cam_high/frame_000547.jpg", "images/human_pika/build_the_blocks/episode_2/cam_high/frame_000240.jpg", "images/human_pika/build_the_blocks/episode_2/cam_left_wrist/frame_000240.jpg", "images/human_pika/build_the_blocks/episode_2/cam_right_wrist/frame_000240.jpg", "images/human_pika/build_the_blocks/episode_2/cam_high/frame_000285.jpg", "images/human_pika/build_the_blocks/episode_2/cam_left_wrist/frame_000285.jpg", "images/human_pika/build_the_blocks/episode_2/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`build the blocks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `build the blocks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.7, "hop_absolute_value": 10.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bread_to_basket_episode_7-train-bf_000045-af_000300-plus-0000395", "task": "pour bread to basket", "image": [ "images/human_pika/pour_bread_to_basket/episode_9/cam_high/frame_000000.jpg", "images/human_pika/pour_bread_to_basket/episode_9/cam_high/frame_000478.jpg", "images/human_pika/pour_bread_to_basket/episode_7/cam_high/frame_000045.jpg", "images/human_pika/pour_bread_to_basket/episode_7/cam_left_wrist/frame_000045.jpg", "images/human_pika/pour_bread_to_basket/episode_7/cam_right_wrist/frame_000045.jpg", "images/human_pika/pour_bread_to_basket/episode_7/cam_high/frame_000300.jpg", "images/human_pika/pour_bread_to_basket/episode_7/cam_left_wrist/frame_000300.jpg", "images/human_pika/pour_bread_to_basket/episode_7/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bread to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bread to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 54.8, "hop_absolute_value": 54.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_banana_episode_24-train-bf_000000-af_000420-plus-0000489", "task": "pour basket with banana", "image": [ "images/human_pika/pour_basket_with_banana/episode_17/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_banana/episode_17/cam_high/frame_000591.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_left_wrist/frame_000000.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_right_wrist/frame_000000.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_high/frame_000420.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_left_wrist/frame_000420.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_right_wrist/frame_000420.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with banana`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with banana`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 82.4, "hop_absolute_value": 82.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_stir_coffee_episode_10-train-bf_000375-af_000495-plus-0000423", "task": "stir coffee", "image": [ "images/human_pika/stir_coffee/episode_23/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_23/cam_high/frame_000575.jpg", "images/human_pika/stir_coffee/episode_10/cam_high/frame_000375.jpg", "images/human_pika/stir_coffee/episode_10/cam_left_wrist/frame_000375.jpg", "images/human_pika/stir_coffee/episode_10/cam_right_wrist/frame_000375.jpg", "images/human_pika/stir_coffee/episode_10/cam_high/frame_000495.jpg", "images/human_pika/stir_coffee/episode_10/cam_left_wrist/frame_000495.jpg", "images/human_pika/stir_coffee/episode_10/cam_right_wrist/frame_000495.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stir coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stir coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_8-train-bf_000405-af_000435-plus-0000789", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_12/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_12/cam_high/frame_000578.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_8/cam_high/frame_000405.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_8/cam_left_wrist/frame_000405.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_8/cam_right_wrist/frame_000405.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_8/cam_high/frame_000435.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_8/cam_left_wrist/frame_000435.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_8/cam_right_wrist/frame_000435.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 22.2, "hop_absolute_value": 22.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_plate_storage_episode_1-train-bf_000030-af_000195-plus-0000814", "task": "pour plate storage", "image": [ "images/human_pika/pour_plate_storage/episode_4/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_4/cam_high/frame_000481.jpg", "images/human_pika/pour_plate_storage/episode_1/cam_high/frame_000030.jpg", "images/human_pika/pour_plate_storage/episode_1/cam_left_wrist/frame_000030.jpg", "images/human_pika/pour_plate_storage/episode_1/cam_right_wrist/frame_000030.jpg", "images/human_pika/pour_plate_storage/episode_1/cam_high/frame_000195.jpg", "images/human_pika/pour_plate_storage/episode_1/cam_left_wrist/frame_000195.jpg", "images/human_pika/pour_plate_storage/episode_1/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour plate storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour plate storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 35.5, "hop_absolute_value": 35.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_3-train-bf_000405-af_000495-plus-0000409", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_high/frame_000543.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_high/frame_000405.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_left_wrist/frame_000405.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_right_wrist/frame_000405.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_high/frame_000495.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_left_wrist/frame_000495.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_right_wrist/frame_000495.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_stack_basket_episode_16-train-bf_000255-af_000690-plus-0000452", "task": "stack basket", "image": [ "images/human_pika/stack_basket/episode_14/cam_high/frame_000000.jpg", "images/human_pika/stack_basket/episode_14/cam_high/frame_000870.jpg", "images/human_pika/stack_basket/episode_16/cam_high/frame_000255.jpg", "images/human_pika/stack_basket/episode_16/cam_left_wrist/frame_000255.jpg", "images/human_pika/stack_basket/episode_16/cam_right_wrist/frame_000255.jpg", "images/human_pika/stack_basket/episode_16/cam_high/frame_000690.jpg", "images/human_pika/stack_basket/episode_16/cam_left_wrist/frame_000690.jpg", "images/human_pika/stack_basket/episode_16/cam_right_wrist/frame_000690.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stack basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stack basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.4, "hop_absolute_value": 60.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_scoop_bean_episode_4-train-bf_000015-af_000465-plus-0000524", "task": "scoop bean", "image": [ "images/human_pika/scoop_bean/episode_2/cam_high/frame_000000.jpg", "images/human_pika/scoop_bean/episode_2/cam_high/frame_000580.jpg", "images/human_pika/scoop_bean/episode_4/cam_high/frame_000015.jpg", "images/human_pika/scoop_bean/episode_4/cam_left_wrist/frame_000015.jpg", "images/human_pika/scoop_bean/episode_4/cam_right_wrist/frame_000015.jpg", "images/human_pika/scoop_bean/episode_4/cam_high/frame_000465.jpg", "images/human_pika/scoop_bean/episode_4/cam_left_wrist/frame_000465.jpg", "images/human_pika/scoop_bean/episode_4/cam_right_wrist/frame_000465.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`scoop bean`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `scoop bean`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 88.2, "hop_absolute_value": 88.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_seal_bag_episode_3-train-bf_000090-af_000255-plus-0000918", "task": "seal bag", "image": [ "images/human_pika/seal_bag/episode_12/cam_high/frame_000000.jpg", "images/human_pika/seal_bag/episode_12/cam_high/frame_000588.jpg", "images/human_pika/seal_bag/episode_3/cam_high/frame_000090.jpg", "images/human_pika/seal_bag/episode_3/cam_left_wrist/frame_000090.jpg", "images/human_pika/seal_bag/episode_3/cam_right_wrist/frame_000090.jpg", "images/human_pika/seal_bag/episode_3/cam_high/frame_000255.jpg", "images/human_pika/seal_bag/episode_3/cam_left_wrist/frame_000255.jpg", "images/human_pika/seal_bag/episode_3/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`seal bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `seal bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 37.9, "hop_absolute_value": 37.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_orange_storage_episode_16-train-bf_000150-af_000225-plus-0000767", "task": "orange storage", "image": [ "images/human_pika/orange_storage/episode_20/cam_high/frame_000000.jpg", "images/human_pika/orange_storage/episode_20/cam_high/frame_000546.jpg", "images/human_pika/orange_storage/episode_16/cam_high/frame_000150.jpg", "images/human_pika/orange_storage/episode_16/cam_left_wrist/frame_000150.jpg", "images/human_pika/orange_storage/episode_16/cam_right_wrist/frame_000150.jpg", "images/human_pika/orange_storage/episode_16/cam_high/frame_000225.jpg", "images/human_pika/orange_storage/episode_16/cam_left_wrist/frame_000225.jpg", "images/human_pika/orange_storage/episode_16/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`orange storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `orange storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 18.5, "hop_absolute_value": 18.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_fill_basket_with_peach_episode_24-train-bf_000405-af_000450-plus-0000875", "task": "fill basket with peach", "image": [ "images/human_pika/fill_basket_with_peach/episode_0/cam_high/frame_000000.jpg", "images/human_pika/fill_basket_with_peach/episode_0/cam_high/frame_000603.jpg", "images/human_pika/fill_basket_with_peach/episode_24/cam_high/frame_000405.jpg", "images/human_pika/fill_basket_with_peach/episode_24/cam_left_wrist/frame_000405.jpg", "images/human_pika/fill_basket_with_peach/episode_24/cam_right_wrist/frame_000405.jpg", "images/human_pika/fill_basket_with_peach/episode_24/cam_high/frame_000450.jpg", "images/human_pika/fill_basket_with_peach/episode_24/cam_left_wrist/frame_000450.jpg", "images/human_pika/fill_basket_with_peach/episode_24/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fill basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fill basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_wipe_stains_episode_12-train-bf_000120-af_000300-plus-0000382", "task": "wipe stains", "image": [ "images/human_pika/wipe_stains/episode_22/cam_high/frame_000000.jpg", "images/human_pika/wipe_stains/episode_22/cam_high/frame_000551.jpg", "images/human_pika/wipe_stains/episode_12/cam_high/frame_000120.jpg", "images/human_pika/wipe_stains/episode_12/cam_left_wrist/frame_000120.jpg", "images/human_pika/wipe_stains/episode_12/cam_right_wrist/frame_000120.jpg", "images/human_pika/wipe_stains/episode_12/cam_high/frame_000300.jpg", "images/human_pika/wipe_stains/episode_12/cam_left_wrist/frame_000300.jpg", "images/human_pika/wipe_stains/episode_12/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`wipe stains`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `wipe stains`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 52.2, "hop_absolute_value": 52.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_fill_basket_with_peach_episode_15-train-bf_000285-af_000315-plus-0000720", "task": "fill basket with peach", "image": [ "images/human_pika/fill_basket_with_peach/episode_22/cam_high/frame_000000.jpg", "images/human_pika/fill_basket_with_peach/episode_22/cam_high/frame_000498.jpg", "images/human_pika/fill_basket_with_peach/episode_15/cam_high/frame_000285.jpg", "images/human_pika/fill_basket_with_peach/episode_15/cam_left_wrist/frame_000285.jpg", "images/human_pika/fill_basket_with_peach/episode_15/cam_right_wrist/frame_000285.jpg", "images/human_pika/fill_basket_with_peach/episode_15/cam_high/frame_000315.jpg", "images/human_pika/fill_basket_with_peach/episode_15/cam_left_wrist/frame_000315.jpg", "images/human_pika/fill_basket_with_peach/episode_15/cam_right_wrist/frame_000315.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fill basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fill basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_fill_basket_with_peach_episode_8-train-bf_000150-af_000270-plus-0000827", "task": "fill basket with peach", "image": [ "images/human_pika/fill_basket_with_peach/episode_10/cam_high/frame_000000.jpg", "images/human_pika/fill_basket_with_peach/episode_10/cam_high/frame_000540.jpg", "images/human_pika/fill_basket_with_peach/episode_8/cam_high/frame_000150.jpg", "images/human_pika/fill_basket_with_peach/episode_8/cam_left_wrist/frame_000150.jpg", "images/human_pika/fill_basket_with_peach/episode_8/cam_right_wrist/frame_000150.jpg", "images/human_pika/fill_basket_with_peach/episode_8/cam_high/frame_000270.jpg", "images/human_pika/fill_basket_with_peach/episode_8/cam_left_wrist/frame_000270.jpg", "images/human_pika/fill_basket_with_peach/episode_8/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fill basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fill basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_plate_storage_episode_9-train-bf_000195-af_000360-plus-0000943", "task": "pour plate storage", "image": [ "images/human_pika/pour_plate_storage/episode_19/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_19/cam_high/frame_000557.jpg", "images/human_pika/pour_plate_storage/episode_9/cam_high/frame_000195.jpg", "images/human_pika/pour_plate_storage/episode_9/cam_left_wrist/frame_000195.jpg", "images/human_pika/pour_plate_storage/episode_9/cam_right_wrist/frame_000195.jpg", "images/human_pika/pour_plate_storage/episode_9/cam_high/frame_000360.jpg", "images/human_pika/pour_plate_storage/episode_9/cam_left_wrist/frame_000360.jpg", "images/human_pika/pour_plate_storage/episode_9/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour plate storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour plate storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 55.0, "hop_absolute_value": 55.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_peach_episode_15-train-bf_000045-af_000465-plus-0000517", "task": "pour basket with peach", "image": [ "images/human_pika/pour_basket_with_peach/episode_15/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_high/frame_000537.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_high/frame_000045.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_left_wrist/frame_000045.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_right_wrist/frame_000045.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_high/frame_000465.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_left_wrist/frame_000465.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_right_wrist/frame_000465.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_scoop_bean_episode_16-train-bf_000120-af_000210-plus-0000752", "task": "scoop bean", "image": [ "images/human_pika/scoop_bean/episode_3/cam_high/frame_000000.jpg", "images/human_pika/scoop_bean/episode_3/cam_high/frame_000590.jpg", "images/human_pika/scoop_bean/episode_16/cam_high/frame_000120.jpg", "images/human_pika/scoop_bean/episode_16/cam_left_wrist/frame_000120.jpg", "images/human_pika/scoop_bean/episode_16/cam_right_wrist/frame_000120.jpg", "images/human_pika/scoop_bean/episode_16/cam_high/frame_000210.jpg", "images/human_pika/scoop_bean/episode_16/cam_left_wrist/frame_000210.jpg", "images/human_pika/scoop_bean/episode_16/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`scoop bean`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `scoop bean`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 24.0, "hop_absolute_value": 24.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_groceries_into_bag_episode_24-train-bf_000330-af_000405-plus-0000914", "task": "put groceries into bag", "image": [ "images/human_pika/put_groceries_into_bag/episode_13/cam_high/frame_000000.jpg", "images/human_pika/put_groceries_into_bag/episode_13/cam_high/frame_000455.jpg", "images/human_pika/put_groceries_into_bag/episode_24/cam_high/frame_000330.jpg", "images/human_pika/put_groceries_into_bag/episode_24/cam_left_wrist/frame_000330.jpg", "images/human_pika/put_groceries_into_bag/episode_24/cam_right_wrist/frame_000330.jpg", "images/human_pika/put_groceries_into_bag/episode_24/cam_high/frame_000405.jpg", "images/human_pika/put_groceries_into_bag/episode_24/cam_left_wrist/frame_000405.jpg", "images/human_pika/put_groceries_into_bag/episode_24/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put groceries into bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put groceries into bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 45.5, "hop_absolute_value": 45.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_22-train-bf_000375-af_000465-plus-0000432", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_22/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_22/cam_high/frame_000502.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_22/cam_high/frame_000375.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_22/cam_left_wrist/frame_000375.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_22/cam_right_wrist/frame_000375.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_22/cam_high/frame_000465.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_22/cam_left_wrist/frame_000465.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_22/cam_right_wrist/frame_000465.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_groceries_into_bag_episode_23-train-bf_000015-af_000375-plus-0000469", "task": "put groceries into bag", "image": [ "images/human_pika/put_groceries_into_bag/episode_2/cam_high/frame_000000.jpg", "images/human_pika/put_groceries_into_bag/episode_2/cam_high/frame_000494.jpg", "images/human_pika/put_groceries_into_bag/episode_23/cam_high/frame_000015.jpg", "images/human_pika/put_groceries_into_bag/episode_23/cam_left_wrist/frame_000015.jpg", "images/human_pika/put_groceries_into_bag/episode_23/cam_right_wrist/frame_000015.jpg", "images/human_pika/put_groceries_into_bag/episode_23/cam_high/frame_000375.jpg", "images/human_pika/put_groceries_into_bag/episode_23/cam_left_wrist/frame_000375.jpg", "images/human_pika/put_groceries_into_bag/episode_23/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put groceries into bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put groceries into bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 82.8, "hop_absolute_value": 82.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_seal_bag_episode_23-train-bf_000270-af_000480-plus-0000533", "task": "seal bag", "image": [ "images/human_pika/seal_bag/episode_21/cam_high/frame_000000.jpg", "images/human_pika/seal_bag/episode_21/cam_high/frame_000542.jpg", "images/human_pika/seal_bag/episode_23/cam_high/frame_000270.jpg", "images/human_pika/seal_bag/episode_23/cam_left_wrist/frame_000270.jpg", "images/human_pika/seal_bag/episode_23/cam_right_wrist/frame_000270.jpg", "images/human_pika/seal_bag/episode_23/cam_high/frame_000480.jpg", "images/human_pika/seal_bag/episode_23/cam_left_wrist/frame_000480.jpg", "images/human_pika/seal_bag/episode_23/cam_right_wrist/frame_000480.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`seal bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `seal bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 93.3, "hop_absolute_value": 93.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_fill_basket_with_peach_episode_13-train-bf_000165-af_000210-plus-0000301", "task": "fill basket with peach", "image": [ "images/human_pika/fill_basket_with_peach/episode_13/cam_high/frame_000000.jpg", "images/human_pika/fill_basket_with_peach/episode_13/cam_high/frame_000542.jpg", "images/human_pika/fill_basket_with_peach/episode_13/cam_high/frame_000165.jpg", "images/human_pika/fill_basket_with_peach/episode_13/cam_left_wrist/frame_000165.jpg", "images/human_pika/fill_basket_with_peach/episode_13/cam_right_wrist/frame_000165.jpg", "images/human_pika/fill_basket_with_peach/episode_13/cam_high/frame_000210.jpg", "images/human_pika/fill_basket_with_peach/episode_13/cam_left_wrist/frame_000210.jpg", "images/human_pika/fill_basket_with_peach/episode_13/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fill basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fill basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.0, "hop_absolute_value": 12.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_seal_bag_episode_9-train-bf_000360-af_000525-plus-0000480", "task": "seal bag", "image": [ "images/human_pika/seal_bag/episode_1/cam_high/frame_000000.jpg", "images/human_pika/seal_bag/episode_1/cam_high/frame_000713.jpg", "images/human_pika/seal_bag/episode_9/cam_high/frame_000360.jpg", "images/human_pika/seal_bag/episode_9/cam_left_wrist/frame_000360.jpg", "images/human_pika/seal_bag/episode_9/cam_right_wrist/frame_000360.jpg", "images/human_pika/seal_bag/episode_9/cam_high/frame_000525.jpg", "images/human_pika/seal_bag/episode_9/cam_left_wrist/frame_000525.jpg", "images/human_pika/seal_bag/episode_9/cam_right_wrist/frame_000525.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`seal bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `seal bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 78.6, "hop_absolute_value": 78.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bread_to_basket_episode_19-train-bf_000120-af_000500-plus-0000555", "task": "pour bread to basket", "image": [ "images/human_pika/pour_bread_to_basket/episode_19/cam_high/frame_000000.jpg", "images/human_pika/pour_bread_to_basket/episode_19/cam_high/frame_000500.jpg", "images/human_pika/pour_bread_to_basket/episode_19/cam_high/frame_000120.jpg", "images/human_pika/pour_bread_to_basket/episode_19/cam_left_wrist/frame_000120.jpg", "images/human_pika/pour_bread_to_basket/episode_19/cam_right_wrist/frame_000120.jpg", "images/human_pika/pour_bread_to_basket/episode_19/cam_high/frame_000500.jpg", "images/human_pika/pour_bread_to_basket/episode_19/cam_left_wrist/frame_000500.jpg", "images/human_pika/pour_bread_to_basket/episode_19/cam_right_wrist/frame_000500.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bread to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bread to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_wipe_stains_episode_6-train-bf_000135-af_000180-plus-0000714", "task": "wipe stains", "image": [ "images/human_pika/wipe_stains/episode_5/cam_high/frame_000000.jpg", "images/human_pika/wipe_stains/episode_5/cam_high/frame_000538.jpg", "images/human_pika/wipe_stains/episode_6/cam_high/frame_000135.jpg", "images/human_pika/wipe_stains/episode_6/cam_left_wrist/frame_000135.jpg", "images/human_pika/wipe_stains/episode_6/cam_right_wrist/frame_000135.jpg", "images/human_pika/wipe_stains/episode_6/cam_high/frame_000180.jpg", "images/human_pika/wipe_stains/episode_6/cam_left_wrist/frame_000180.jpg", "images/human_pika/wipe_stains/episode_6/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`wipe stains`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `wipe stains`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_banana_episode_12-train-bf_000345-af_000435-plus-0000353", "task": "pour basket with banana", "image": [ "images/human_pika/pour_basket_with_banana/episode_0/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_banana/episode_0/cam_high/frame_000558.jpg", "images/human_pika/pour_basket_with_banana/episode_12/cam_high/frame_000345.jpg", "images/human_pika/pour_basket_with_banana/episode_12/cam_left_wrist/frame_000345.jpg", "images/human_pika/pour_basket_with_banana/episode_12/cam_right_wrist/frame_000345.jpg", "images/human_pika/pour_basket_with_banana/episode_12/cam_high/frame_000435.jpg", "images/human_pika/pour_basket_with_banana/episode_12/cam_left_wrist/frame_000435.jpg", "images/human_pika/pour_basket_with_banana/episode_12/cam_right_wrist/frame_000435.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with banana`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with banana`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 37.5, "hop_absolute_value": 37.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bread_to_basket_episode_0-train-bf_000015-af_000075-plus-0000647", "task": "pour bread to basket", "image": [ "images/human_pika/pour_bread_to_basket/episode_17/cam_high/frame_000000.jpg", "images/human_pika/pour_bread_to_basket/episode_17/cam_high/frame_000505.jpg", "images/human_pika/pour_bread_to_basket/episode_0/cam_high/frame_000015.jpg", "images/human_pika/pour_bread_to_basket/episode_0/cam_left_wrist/frame_000015.jpg", "images/human_pika/pour_bread_to_basket/episode_0/cam_right_wrist/frame_000015.jpg", "images/human_pika/pour_bread_to_basket/episode_0/cam_high/frame_000075.jpg", "images/human_pika/pour_bread_to_basket/episode_0/cam_left_wrist/frame_000075.jpg", "images/human_pika/pour_bread_to_basket/episode_0/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bread to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bread to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_24-train-bf_000255-af_000510-plus-0000485", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_0/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_0/cam_high/frame_000636.jpg", "images/human_pika/pour_tea/episode_24/cam_high/frame_000255.jpg", "images/human_pika/pour_tea/episode_24/cam_left_wrist/frame_000255.jpg", "images/human_pika/pour_tea/episode_24/cam_right_wrist/frame_000255.jpg", "images/human_pika/pour_tea/episode_24/cam_high/frame_000510.jpg", "images/human_pika/pour_tea/episode_24/cam_left_wrist/frame_000510.jpg", "images/human_pika/pour_tea/episode_24/cam_right_wrist/frame_000510.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 77.3, "hop_absolute_value": 77.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_23-train-bf_000045-af_000150-plus-0000729", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_12/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_12/cam_high/frame_000578.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_high/frame_000045.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_left_wrist/frame_000045.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_right_wrist/frame_000045.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_high/frame_000150.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_left_wrist/frame_000150.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 24.1, "hop_absolute_value": 24.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_peach_episode_16-train-bf_000375-af_000435-plus-0000930", "task": "pour basket with peach", "image": [ "images/human_pika/pour_basket_with_peach/episode_16/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_peach/episode_16/cam_high/frame_000544.jpg", "images/human_pika/pour_basket_with_peach/episode_16/cam_high/frame_000375.jpg", "images/human_pika/pour_basket_with_peach/episode_16/cam_left_wrist/frame_000375.jpg", "images/human_pika/pour_basket_with_peach/episode_16/cam_right_wrist/frame_000375.jpg", "images/human_pika/pour_basket_with_peach/episode_16/cam_high/frame_000435.jpg", "images/human_pika/pour_basket_with_peach/episode_16/cam_left_wrist/frame_000435.jpg", "images/human_pika/pour_basket_with_peach/episode_16/cam_right_wrist/frame_000435.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 36.4, "hop_absolute_value": 36.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_17-train-bf_000015-af_000135-plus-0000817", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_24/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_24/cam_high/frame_000509.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_high/frame_000015.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_left_wrist/frame_000015.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_right_wrist/frame_000015.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_high/frame_000135.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_left_wrist/frame_000135.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 22.2, "hop_absolute_value": 22.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_peach_to_the_basket_episode_2-train-bf_000105-af_000315-plus-0000370", "task": "put peach to the basket", "image": [ "images/human_pika/put_peach_to_the_basket/episode_16/cam_high/frame_000000.jpg", "images/human_pika/put_peach_to_the_basket/episode_16/cam_high/frame_000544.jpg", "images/human_pika/put_peach_to_the_basket/episode_2/cam_high/frame_000105.jpg", "images/human_pika/put_peach_to_the_basket/episode_2/cam_left_wrist/frame_000105.jpg", "images/human_pika/put_peach_to_the_basket/episode_2/cam_right_wrist/frame_000105.jpg", "images/human_pika/put_peach_to_the_basket/episode_2/cam_high/frame_000315.jpg", "images/human_pika/put_peach_to_the_basket/episode_2/cam_left_wrist/frame_000315.jpg", "images/human_pika/put_peach_to_the_basket/episode_2/cam_right_wrist/frame_000315.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put peach to the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put peach to the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 48.3, "hop_absolute_value": 48.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_3-train-bf_000195-af_000255-plus-0000733", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_high/frame_000543.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_high/frame_000195.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_left_wrist/frame_000195.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_right_wrist/frame_000195.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_high/frame_000255.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_left_wrist/frame_000255.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 17.4, "hop_absolute_value": 17.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_peach_episode_18-train-bf_000315-af_000330-plus-0000640", "task": "pour basket with peach", "image": [ "images/human_pika/pour_basket_with_peach/episode_21/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_peach/episode_21/cam_high/frame_000503.jpg", "images/human_pika/pour_basket_with_peach/episode_18/cam_high/frame_000315.jpg", "images/human_pika/pour_basket_with_peach/episode_18/cam_left_wrist/frame_000315.jpg", "images/human_pika/pour_basket_with_peach/episode_18/cam_right_wrist/frame_000315.jpg", "images/human_pika/pour_basket_with_peach/episode_18/cam_high/frame_000330.jpg", "images/human_pika/pour_basket_with_peach/episode_18/cam_left_wrist/frame_000330.jpg", "images/human_pika/pour_basket_with_peach/episode_18/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 6.7, "hop_absolute_value": 6.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_10-train-bf_000000-af_000165-plus-0000931", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_16/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_16/cam_high/frame_000562.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_10/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_10/cam_left_wrist/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_10/cam_right_wrist/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_10/cam_high/frame_000165.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_10/cam_left_wrist/frame_000165.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_10/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 29.7, "hop_absolute_value": 29.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_scoop_bean_episode_23-train-bf_000240-af_000270-plus-0000639", "task": "scoop bean", "image": [ "images/human_pika/scoop_bean/episode_19/cam_high/frame_000000.jpg", "images/human_pika/scoop_bean/episode_19/cam_high/frame_000461.jpg", "images/human_pika/scoop_bean/episode_23/cam_high/frame_000240.jpg", "images/human_pika/scoop_bean/episode_23/cam_left_wrist/frame_000240.jpg", "images/human_pika/scoop_bean/episode_23/cam_right_wrist/frame_000240.jpg", "images/human_pika/scoop_bean/episode_23/cam_high/frame_000270.jpg", "images/human_pika/scoop_bean/episode_23/cam_left_wrist/frame_000270.jpg", "images/human_pika/scoop_bean/episode_23/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`scoop bean`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `scoop bean`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 13.3, "hop_absolute_value": 13.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_scoop_bean_episode_17-train-bf_000180-af_000375-plus-0000376", "task": "scoop bean", "image": [ "images/human_pika/scoop_bean/episode_24/cam_high/frame_000000.jpg", "images/human_pika/scoop_bean/episode_24/cam_high/frame_000544.jpg", "images/human_pika/scoop_bean/episode_17/cam_high/frame_000180.jpg", "images/human_pika/scoop_bean/episode_17/cam_left_wrist/frame_000180.jpg", "images/human_pika/scoop_bean/episode_17/cam_right_wrist/frame_000180.jpg", "images/human_pika/scoop_bean/episode_17/cam_high/frame_000375.jpg", "images/human_pika/scoop_bean/episode_17/cam_left_wrist/frame_000375.jpg", "images/human_pika/scoop_bean/episode_17/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`scoop bean`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `scoop bean`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_peach_to_the_basket_episode_13-train-bf_000015-af_000300-plus-0000404", "task": "put peach to the basket", "image": [ "images/human_pika/put_peach_to_the_basket/episode_4/cam_high/frame_000000.jpg", "images/human_pika/put_peach_to_the_basket/episode_4/cam_high/frame_000534.jpg", "images/human_pika/put_peach_to_the_basket/episode_13/cam_high/frame_000015.jpg", "images/human_pika/put_peach_to_the_basket/episode_13/cam_left_wrist/frame_000015.jpg", "images/human_pika/put_peach_to_the_basket/episode_13/cam_right_wrist/frame_000015.jpg", "images/human_pika/put_peach_to_the_basket/episode_13/cam_high/frame_000300.jpg", "images/human_pika/put_peach_to_the_basket/episode_13/cam_left_wrist/frame_000300.jpg", "images/human_pika/put_peach_to_the_basket/episode_13/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put peach to the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put peach to the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 54.3, "hop_absolute_value": 54.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_groceries_into_bag_episode_7-train-bf_000105-af_000285-plus-0000854", "task": "put groceries into bag", "image": [ "images/human_pika/put_groceries_into_bag/episode_12/cam_high/frame_000000.jpg", "images/human_pika/put_groceries_into_bag/episode_12/cam_high/frame_000492.jpg", "images/human_pika/put_groceries_into_bag/episode_7/cam_high/frame_000105.jpg", "images/human_pika/put_groceries_into_bag/episode_7/cam_left_wrist/frame_000105.jpg", "images/human_pika/put_groceries_into_bag/episode_7/cam_right_wrist/frame_000105.jpg", "images/human_pika/put_groceries_into_bag/episode_7/cam_high/frame_000285.jpg", "images/human_pika/put_groceries_into_bag/episode_7/cam_left_wrist/frame_000285.jpg", "images/human_pika/put_groceries_into_bag/episode_7/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put groceries into bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put groceries into bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 48.0, "hop_absolute_value": 48.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_peach_episode_17-train-bf_000105-af_000180-plus-0000726", "task": "pour basket with peach", "image": [ "images/human_pika/pour_basket_with_peach/episode_7/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_peach/episode_7/cam_high/frame_000521.jpg", "images/human_pika/pour_basket_with_peach/episode_17/cam_high/frame_000105.jpg", "images/human_pika/pour_basket_with_peach/episode_17/cam_left_wrist/frame_000105.jpg", "images/human_pika/pour_basket_with_peach/episode_17/cam_right_wrist/frame_000105.jpg", "images/human_pika/pour_basket_with_peach/episode_17/cam_high/frame_000180.jpg", "images/human_pika/pour_basket_with_peach/episode_17/cam_left_wrist/frame_000180.jpg", "images/human_pika/pour_basket_with_peach/episode_17/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_build_blocks_episode_17-train-bf_000090-af_000285-plus-0000355", "task": "build blocks", "image": [ "images/human_pika/build_blocks/episode_18/cam_high/frame_000000.jpg", "images/human_pika/build_blocks/episode_18/cam_high/frame_000454.jpg", "images/human_pika/build_blocks/episode_17/cam_high/frame_000090.jpg", "images/human_pika/build_blocks/episode_17/cam_left_wrist/frame_000090.jpg", "images/human_pika/build_blocks/episode_17/cam_right_wrist/frame_000090.jpg", "images/human_pika/build_blocks/episode_17/cam_high/frame_000285.jpg", "images/human_pika/build_blocks/episode_17/cam_left_wrist/frame_000285.jpg", "images/human_pika/build_blocks/episode_17/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`build blocks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `build blocks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 43.3, "hop_absolute_value": 43.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_pants_episode_3-train-bf_000315-af_000435-plus-0000384", "task": "fold pants", "image": [ "images/human_pika/fold_pants/episode_0/cam_high/frame_000000.jpg", "images/human_pika/fold_pants/episode_0/cam_high/frame_000680.jpg", "images/human_pika/fold_pants/episode_3/cam_high/frame_000315.jpg", "images/human_pika/fold_pants/episode_3/cam_left_wrist/frame_000315.jpg", "images/human_pika/fold_pants/episode_3/cam_right_wrist/frame_000315.jpg", "images/human_pika/fold_pants/episode_3/cam_high/frame_000435.jpg", "images/human_pika/fold_pants/episode_3/cam_left_wrist/frame_000435.jpg", "images/human_pika/fold_pants/episode_3/cam_right_wrist/frame_000435.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold pants`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold pants`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 53.3, "hop_absolute_value": 53.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_pants_episode_16-train-bf_000120-af_000195-plus-0000298", "task": "fold pants", "image": [ "images/human_pika/fold_pants/episode_22/cam_high/frame_000000.jpg", "images/human_pika/fold_pants/episode_22/cam_high/frame_000508.jpg", "images/human_pika/fold_pants/episode_16/cam_high/frame_000120.jpg", "images/human_pika/fold_pants/episode_16/cam_left_wrist/frame_000120.jpg", "images/human_pika/fold_pants/episode_16/cam_right_wrist/frame_000120.jpg", "images/human_pika/fold_pants/episode_16/cam_high/frame_000195.jpg", "images/human_pika/fold_pants/episode_16/cam_left_wrist/frame_000195.jpg", "images/human_pika/fold_pants/episode_16/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold pants`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold pants`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 17.9, "hop_absolute_value": 17.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_banana_episode_17-train-bf_000360-af_000435-plus-0000333", "task": "pour basket with banana", "image": [ "images/human_pika/pour_basket_with_banana/episode_4/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_banana/episode_4/cam_high/frame_000528.jpg", "images/human_pika/pour_basket_with_banana/episode_17/cam_high/frame_000360.jpg", "images/human_pika/pour_basket_with_banana/episode_17/cam_left_wrist/frame_000360.jpg", "images/human_pika/pour_basket_with_banana/episode_17/cam_right_wrist/frame_000360.jpg", "images/human_pika/pour_basket_with_banana/episode_17/cam_high/frame_000435.jpg", "images/human_pika/pour_basket_with_banana/episode_17/cam_left_wrist/frame_000435.jpg", "images/human_pika/pour_basket_with_banana/episode_17/cam_right_wrist/frame_000435.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with banana`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with banana`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_11-train-bf_000450-af_000480-plus-0000880", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_11/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_11/cam_high/frame_000554.jpg", "images/human_pika/pour_tea/episode_11/cam_high/frame_000450.jpg", "images/human_pika/pour_tea/episode_11/cam_left_wrist/frame_000450.jpg", "images/human_pika/pour_tea/episode_11/cam_right_wrist/frame_000450.jpg", "images/human_pika/pour_tea/episode_11/cam_high/frame_000480.jpg", "images/human_pika/pour_tea/episode_11/cam_left_wrist/frame_000480.jpg", "images/human_pika/pour_tea/episode_11/cam_right_wrist/frame_000480.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_towel_episode_19-train-bf_000135-af_000405-plus-0000481", "task": "fold towel", "image": [ "images/human_pika/fold_towel/episode_19/cam_high/frame_000000.jpg", "images/human_pika/fold_towel/episode_19/cam_high/frame_000449.jpg", "images/human_pika/fold_towel/episode_19/cam_high/frame_000135.jpg", "images/human_pika/fold_towel/episode_19/cam_left_wrist/frame_000135.jpg", "images/human_pika/fold_towel/episode_19/cam_right_wrist/frame_000135.jpg", "images/human_pika/fold_towel/episode_19/cam_high/frame_000405.jpg", "images/human_pika/fold_towel/episode_19/cam_left_wrist/frame_000405.jpg", "images/human_pika/fold_towel/episode_19/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold towel`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold towel`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 90.0, "hop_absolute_value": 90.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_towel_episode_6-train-bf_000360-af_000480-plus-0000507", "task": "fold towel", "image": [ "images/human_pika/fold_towel/episode_6/cam_high/frame_000000.jpg", "images/human_pika/fold_towel/episode_6/cam_high/frame_000480.jpg", "images/human_pika/fold_towel/episode_6/cam_high/frame_000360.jpg", "images/human_pika/fold_towel/episode_6/cam_left_wrist/frame_000360.jpg", "images/human_pika/fold_towel/episode_6/cam_right_wrist/frame_000360.jpg", "images/human_pika/fold_towel/episode_6/cam_high/frame_000480.jpg", "images/human_pika/fold_towel/episode_6/cam_left_wrist/frame_000480.jpg", "images/human_pika/fold_towel/episode_6/cam_right_wrist/frame_000480.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold towel`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold towel`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_stack_basket_episode_13-train-bf_000345-af_000360-plus-0000665", "task": "stack basket", "image": [ "images/human_pika/stack_basket/episode_15/cam_high/frame_000000.jpg", "images/human_pika/stack_basket/episode_15/cam_high/frame_000978.jpg", "images/human_pika/stack_basket/episode_13/cam_high/frame_000345.jpg", "images/human_pika/stack_basket/episode_13/cam_left_wrist/frame_000345.jpg", "images/human_pika/stack_basket/episode_13/cam_right_wrist/frame_000345.jpg", "images/human_pika/stack_basket/episode_13/cam_high/frame_000360.jpg", "images/human_pika/stack_basket/episode_13/cam_left_wrist/frame_000360.jpg", "images/human_pika/stack_basket/episode_13/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stack basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stack basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.7, "hop_absolute_value": 1.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bread_to_basket_episode_2-train-bf_000315-af_000375-plus-0000846", "task": "pour bread to basket", "image": [ "images/human_pika/pour_bread_to_basket/episode_0/cam_high/frame_000000.jpg", "images/human_pika/pour_bread_to_basket/episode_0/cam_high/frame_000562.jpg", "images/human_pika/pour_bread_to_basket/episode_2/cam_high/frame_000315.jpg", "images/human_pika/pour_bread_to_basket/episode_2/cam_left_wrist/frame_000315.jpg", "images/human_pika/pour_bread_to_basket/episode_2/cam_right_wrist/frame_000315.jpg", "images/human_pika/pour_bread_to_basket/episode_2/cam_high/frame_000375.jpg", "images/human_pika/pour_bread_to_basket/episode_2/cam_left_wrist/frame_000375.jpg", "images/human_pika/pour_bread_to_basket/episode_2/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bread to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bread to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 28.6, "hop_absolute_value": 28.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_build_blocks_episode_16-train-bf_000270-af_000450-plus-0000522", "task": "build blocks", "image": [ "images/human_pika/build_blocks/episode_16/cam_high/frame_000000.jpg", "images/human_pika/build_blocks/episode_16/cam_high/frame_000474.jpg", "images/human_pika/build_blocks/episode_16/cam_high/frame_000270.jpg", "images/human_pika/build_blocks/episode_16/cam_left_wrist/frame_000270.jpg", "images/human_pika/build_blocks/episode_16/cam_right_wrist/frame_000270.jpg", "images/human_pika/build_blocks/episode_16/cam_high/frame_000450.jpg", "images/human_pika/build_blocks/episode_16/cam_left_wrist/frame_000450.jpg", "images/human_pika/build_blocks/episode_16/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`build blocks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `build blocks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 92.3, "hop_absolute_value": 92.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_pants_episode_9-train-bf_000105-af_000135-plus-0000591", "task": "fold pants", "image": [ "images/human_pika/fold_pants/episode_9/cam_high/frame_000000.jpg", "images/human_pika/fold_pants/episode_9/cam_high/frame_000468.jpg", "images/human_pika/fold_pants/episode_9/cam_high/frame_000105.jpg", "images/human_pika/fold_pants/episode_9/cam_left_wrist/frame_000105.jpg", "images/human_pika/fold_pants/episode_9/cam_right_wrist/frame_000105.jpg", "images/human_pika/fold_pants/episode_9/cam_high/frame_000135.jpg", "images/human_pika/fold_pants/episode_9/cam_left_wrist/frame_000135.jpg", "images/human_pika/fold_pants/episode_9/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold pants`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold pants`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 8.3, "hop_absolute_value": 8.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_stack_basket_episode_0-train-bf_000270-af_000285-plus-0000651", "task": "stack basket", "image": [ "images/human_pika/stack_basket/episode_24/cam_high/frame_000000.jpg", "images/human_pika/stack_basket/episode_24/cam_high/frame_000623.jpg", "images/human_pika/stack_basket/episode_0/cam_high/frame_000270.jpg", "images/human_pika/stack_basket/episode_0/cam_left_wrist/frame_000270.jpg", "images/human_pika/stack_basket/episode_0/cam_right_wrist/frame_000270.jpg", "images/human_pika/stack_basket/episode_0/cam_high/frame_000285.jpg", "images/human_pika/stack_basket/episode_0/cam_left_wrist/frame_000285.jpg", "images/human_pika/stack_basket/episode_0/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stack basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stack basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.4, "hop_absolute_value": 2.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_stack_basket_episode_13-train-bf_000825-af_001020-plus-0000397", "task": "stack basket", "image": [ "images/human_pika/stack_basket/episode_13/cam_high/frame_000000.jpg", "images/human_pika/stack_basket/episode_13/cam_high/frame_001248.jpg", "images/human_pika/stack_basket/episode_13/cam_high/frame_000825.jpg", "images/human_pika/stack_basket/episode_13/cam_left_wrist/frame_000825.jpg", "images/human_pika/stack_basket/episode_13/cam_right_wrist/frame_000825.jpg", "images/human_pika/stack_basket/episode_13/cam_high/frame_001020.jpg", "images/human_pika/stack_basket/episode_13/cam_left_wrist/frame_001020.jpg", "images/human_pika/stack_basket/episode_13/cam_right_wrist/frame_001020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stack basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stack basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 46.4, "hop_absolute_value": 46.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_seal_bag_episode_23-train-bf_000105-af_000255-plus-0000347", "task": "seal bag", "image": [ "images/human_pika/seal_bag/episode_23/cam_high/frame_000000.jpg", "images/human_pika/seal_bag/episode_23/cam_high/frame_000506.jpg", "images/human_pika/seal_bag/episode_23/cam_high/frame_000105.jpg", "images/human_pika/seal_bag/episode_23/cam_left_wrist/frame_000105.jpg", "images/human_pika/seal_bag/episode_23/cam_right_wrist/frame_000105.jpg", "images/human_pika/seal_bag/episode_23/cam_high/frame_000255.jpg", "images/human_pika/seal_bag/episode_23/cam_left_wrist/frame_000255.jpg", "images/human_pika/seal_bag/episode_23/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`seal bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `seal bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 38.5, "hop_absolute_value": 38.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_build_blocks_episode_19-train-bf_000240-af_000330-plus-0000332", "task": "build blocks", "image": [ "images/human_pika/build_blocks/episode_1/cam_high/frame_000000.jpg", "images/human_pika/build_blocks/episode_1/cam_high/frame_000488.jpg", "images/human_pika/build_blocks/episode_19/cam_high/frame_000240.jpg", "images/human_pika/build_blocks/episode_19/cam_left_wrist/frame_000240.jpg", "images/human_pika/build_blocks/episode_19/cam_right_wrist/frame_000240.jpg", "images/human_pika/build_blocks/episode_19/cam_high/frame_000330.jpg", "images/human_pika/build_blocks/episode_19/cam_left_wrist/frame_000330.jpg", "images/human_pika/build_blocks/episode_19/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`build blocks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `build blocks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 37.5, "hop_absolute_value": 37.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_towel_episode_4-train-bf_000135-af_000450-plus-0000501", "task": "fold towel", "image": [ "images/human_pika/fold_towel/episode_21/cam_high/frame_000000.jpg", "images/human_pika/fold_towel/episode_21/cam_high/frame_000495.jpg", "images/human_pika/fold_towel/episode_4/cam_high/frame_000135.jpg", "images/human_pika/fold_towel/episode_4/cam_left_wrist/frame_000135.jpg", "images/human_pika/fold_towel/episode_4/cam_right_wrist/frame_000135.jpg", "images/human_pika/fold_towel/episode_4/cam_high/frame_000450.jpg", "images/human_pika/fold_towel/episode_4/cam_left_wrist/frame_000450.jpg", "images/human_pika/fold_towel/episode_4/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold towel`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold towel`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_pants_episode_18-train-bf_000000-af_000105-plus-0000736", "task": "fold pants", "image": [ "images/human_pika/fold_pants/episode_18/cam_high/frame_000000.jpg", "images/human_pika/fold_pants/episode_18/cam_high/frame_000499.jpg", "images/human_pika/fold_pants/episode_18/cam_high/frame_000000.jpg", "images/human_pika/fold_pants/episode_18/cam_left_wrist/frame_000000.jpg", "images/human_pika/fold_pants/episode_18/cam_right_wrist/frame_000000.jpg", "images/human_pika/fold_pants/episode_18/cam_high/frame_000105.jpg", "images/human_pika/fold_pants/episode_18/cam_left_wrist/frame_000105.jpg", "images/human_pika/fold_pants/episode_18/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold pants`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold pants`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 21.2, "hop_absolute_value": 21.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_plate_storage_episode_13-train-bf_000150-af_000195-plus-0000666", "task": "pour plate storage", "image": [ "images/human_pika/pour_plate_storage/episode_13/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_13/cam_high/frame_000491.jpg", "images/human_pika/pour_plate_storage/episode_13/cam_high/frame_000150.jpg", "images/human_pika/pour_plate_storage/episode_13/cam_left_wrist/frame_000150.jpg", "images/human_pika/pour_plate_storage/episode_13/cam_right_wrist/frame_000150.jpg", "images/human_pika/pour_plate_storage/episode_13/cam_high/frame_000195.jpg", "images/human_pika/pour_plate_storage/episode_13/cam_left_wrist/frame_000195.jpg", "images/human_pika/pour_plate_storage/episode_13/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour plate storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour plate storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 13.6, "hop_absolute_value": 13.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_pants_episode_4-train-bf_000075-af_000150-plus-0000766", "task": "fold pants", "image": [ "images/human_pika/fold_pants/episode_2/cam_high/frame_000000.jpg", "images/human_pika/fold_pants/episode_2/cam_high/frame_000474.jpg", "images/human_pika/fold_pants/episode_4/cam_high/frame_000075.jpg", "images/human_pika/fold_pants/episode_4/cam_left_wrist/frame_000075.jpg", "images/human_pika/fold_pants/episode_4/cam_right_wrist/frame_000075.jpg", "images/human_pika/fold_pants/episode_4/cam_high/frame_000150.jpg", "images/human_pika/fold_pants/episode_4/cam_left_wrist/frame_000150.jpg", "images/human_pika/fold_pants/episode_4/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold pants`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold pants`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 15.6, "hop_absolute_value": 15.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_test_tube_to_the_shelf_episode_0-train-bf_000060-af_000460-plus-0000922", "task": "put test tube to the shelf", "image": [ "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_high/frame_000000.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_high/frame_000460.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_high/frame_000060.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_left_wrist/frame_000060.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_right_wrist/frame_000060.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_high/frame_000460.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_left_wrist/frame_000460.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_right_wrist/frame_000460.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put test tube to the shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put test tube to the shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_build_the_blocks_episode_13-train-bf_000195-af_000511-plus-0000544", "task": "build the blocks", "image": [ "images/human_pika/build_the_blocks/episode_20/cam_high/frame_000000.jpg", "images/human_pika/build_the_blocks/episode_20/cam_high/frame_000541.jpg", "images/human_pika/build_the_blocks/episode_13/cam_high/frame_000195.jpg", "images/human_pika/build_the_blocks/episode_13/cam_left_wrist/frame_000195.jpg", "images/human_pika/build_the_blocks/episode_13/cam_right_wrist/frame_000195.jpg", "images/human_pika/build_the_blocks/episode_13/cam_high/frame_000511.jpg", "images/human_pika/build_the_blocks/episode_13/cam_left_wrist/frame_000511.jpg", "images/human_pika/build_the_blocks/episode_13/cam_right_wrist/frame_000511.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`build the blocks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `build the blocks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_place_test_tube_episode_22-train-bf_000345-af_000360-plus-0000661", "task": "place test tube", "image": [ "images/human_pika/place_test_tube/episode_16/cam_high/frame_000000.jpg", "images/human_pika/place_test_tube/episode_16/cam_high/frame_000559.jpg", "images/human_pika/place_test_tube/episode_22/cam_high/frame_000345.jpg", "images/human_pika/place_test_tube/episode_22/cam_left_wrist/frame_000345.jpg", "images/human_pika/place_test_tube/episode_22/cam_right_wrist/frame_000345.jpg", "images/human_pika/place_test_tube/episode_22/cam_high/frame_000360.jpg", "images/human_pika/place_test_tube/episode_22/cam_left_wrist/frame_000360.jpg", "images/human_pika/place_test_tube/episode_22/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`place test tube`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `place test tube`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_wipe_stains_episode_19-train-bf_000105-af_000120-plus-0000582", "task": "wipe stains", "image": [ "images/human_pika/wipe_stains/episode_11/cam_high/frame_000000.jpg", "images/human_pika/wipe_stains/episode_11/cam_high/frame_000476.jpg", "images/human_pika/wipe_stains/episode_19/cam_high/frame_000105.jpg", "images/human_pika/wipe_stains/episode_19/cam_left_wrist/frame_000105.jpg", "images/human_pika/wipe_stains/episode_19/cam_right_wrist/frame_000105.jpg", "images/human_pika/wipe_stains/episode_19/cam_high/frame_000120.jpg", "images/human_pika/wipe_stains/episode_19/cam_left_wrist/frame_000120.jpg", "images/human_pika/wipe_stains/episode_19/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`wipe stains`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `wipe stains`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.0, "hop_absolute_value": 3.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_seal_bag_episode_1-train-bf_000075-af_000090-plus-0000611", "task": "seal bag", "image": [ "images/human_pika/seal_bag/episode_16/cam_high/frame_000000.jpg", "images/human_pika/seal_bag/episode_16/cam_high/frame_000589.jpg", "images/human_pika/seal_bag/episode_1/cam_high/frame_000075.jpg", "images/human_pika/seal_bag/episode_1/cam_left_wrist/frame_000075.jpg", "images/human_pika/seal_bag/episode_1/cam_right_wrist/frame_000075.jpg", "images/human_pika/seal_bag/episode_1/cam_high/frame_000090.jpg", "images/human_pika/seal_bag/episode_1/cam_left_wrist/frame_000090.jpg", "images/human_pika/seal_bag/episode_1/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`seal bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `seal bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.4, "hop_absolute_value": 2.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_peach_to_the_basket_episode_16-train-bf_000000-af_000015-plus-0000568", "task": "put peach to the basket", "image": [ "images/human_pika/put_peach_to_the_basket/episode_16/cam_high/frame_000000.jpg", "images/human_pika/put_peach_to_the_basket/episode_16/cam_high/frame_000544.jpg", "images/human_pika/put_peach_to_the_basket/episode_16/cam_high/frame_000000.jpg", "images/human_pika/put_peach_to_the_basket/episode_16/cam_left_wrist/frame_000000.jpg", "images/human_pika/put_peach_to_the_basket/episode_16/cam_right_wrist/frame_000000.jpg", "images/human_pika/put_peach_to_the_basket/episode_16/cam_high/frame_000015.jpg", "images/human_pika/put_peach_to_the_basket/episode_16/cam_left_wrist/frame_000015.jpg", "images/human_pika/put_peach_to_the_basket/episode_16/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put peach to the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put peach to the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.8, "hop_absolute_value": 2.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_11-train-bf_000075-af_000090-plus-0000564", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_11/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_11/cam_high/frame_000554.jpg", "images/human_pika/pour_tea/episode_11/cam_high/frame_000075.jpg", "images/human_pika/pour_tea/episode_11/cam_left_wrist/frame_000075.jpg", "images/human_pika/pour_tea/episode_11/cam_right_wrist/frame_000075.jpg", "images/human_pika/pour_tea/episode_11/cam_high/frame_000090.jpg", "images/human_pika/pour_tea/episode_11/cam_left_wrist/frame_000090.jpg", "images/human_pika/pour_tea/episode_11/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.2, "hop_absolute_value": 3.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_orange_storage_episode_13-train-bf_000255-af_000270-plus-0000657", "task": "orange storage", "image": [ "images/human_pika/orange_storage/episode_13/cam_high/frame_000000.jpg", "images/human_pika/orange_storage/episode_13/cam_high/frame_000612.jpg", "images/human_pika/orange_storage/episode_13/cam_high/frame_000255.jpg", "images/human_pika/orange_storage/episode_13/cam_left_wrist/frame_000255.jpg", "images/human_pika/orange_storage/episode_13/cam_right_wrist/frame_000255.jpg", "images/human_pika/orange_storage/episode_13/cam_high/frame_000270.jpg", "images/human_pika/orange_storage/episode_13/cam_left_wrist/frame_000270.jpg", "images/human_pika/orange_storage/episode_13/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`orange storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `orange storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.3, "hop_absolute_value": 4.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_wipe_stains_episode_18-train-bf_000225-af_000105-minus-0000149", "task": "wipe stains", "image": [ "images/human_pika/wipe_stains/episode_1/cam_high/frame_000000.jpg", "images/human_pika/wipe_stains/episode_1/cam_high/frame_000609.jpg", "images/human_pika/wipe_stains/episode_18/cam_high/frame_000225.jpg", "images/human_pika/wipe_stains/episode_18/cam_left_wrist/frame_000225.jpg", "images/human_pika/wipe_stains/episode_18/cam_right_wrist/frame_000225.jpg", "images/human_pika/wipe_stains/episode_18/cam_high/frame_000105.jpg", "images/human_pika/wipe_stains/episode_18/cam_left_wrist/frame_000105.jpg", "images/human_pika/wipe_stains/episode_18/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`wipe stains`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `wipe stains`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -53.3, "hop_absolute_value": 53.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_plate_storage_episode_11-train-bf_000420-af_000090-minus-0000088", "task": "pour plate storage", "image": [ "images/human_pika/pour_plate_storage/episode_13/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_13/cam_high/frame_000491.jpg", "images/human_pika/pour_plate_storage/episode_11/cam_high/frame_000420.jpg", "images/human_pika/pour_plate_storage/episode_11/cam_left_wrist/frame_000420.jpg", "images/human_pika/pour_plate_storage/episode_11/cam_right_wrist/frame_000420.jpg", "images/human_pika/pour_plate_storage/episode_11/cam_high/frame_000090.jpg", "images/human_pika/pour_plate_storage/episode_11/cam_left_wrist/frame_000090.jpg", "images/human_pika/pour_plate_storage/episode_11/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour plate storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour plate storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -78.6, "hop_absolute_value": 78.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_scoop_bean_episode_10-train-bf_000150-af_000090-minus-0000800", "task": "scoop bean", "image": [ "images/human_pika/scoop_bean/episode_23/cam_high/frame_000000.jpg", "images/human_pika/scoop_bean/episode_23/cam_high/frame_000476.jpg", "images/human_pika/scoop_bean/episode_10/cam_high/frame_000150.jpg", "images/human_pika/scoop_bean/episode_10/cam_left_wrist/frame_000150.jpg", "images/human_pika/scoop_bean/episode_10/cam_right_wrist/frame_000150.jpg", "images/human_pika/scoop_bean/episode_10/cam_high/frame_000090.jpg", "images/human_pika/scoop_bean/episode_10/cam_left_wrist/frame_000090.jpg", "images/human_pika/scoop_bean/episode_10/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`scoop bean`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `scoop bean`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_banana_episode_24-train-bf_000270-af_000015-minus-0000016", "task": "pour basket with banana", "image": [ "images/human_pika/pour_basket_with_banana/episode_24/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_high/frame_000513.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_high/frame_000270.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_left_wrist/frame_000270.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_right_wrist/frame_000270.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_high/frame_000015.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_left_wrist/frame_000015.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with banana`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with banana`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -94.4, "hop_absolute_value": 94.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_stir_coffee_episode_0-train-bf_000045-af_000015-minus-0000124", "task": "stir coffee", "image": [ "images/human_pika/stir_coffee/episode_20/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_20/cam_high/frame_000605.jpg", "images/human_pika/stir_coffee/episode_0/cam_high/frame_000045.jpg", "images/human_pika/stir_coffee/episode_0/cam_left_wrist/frame_000045.jpg", "images/human_pika/stir_coffee/episode_0/cam_right_wrist/frame_000045.jpg", "images/human_pika/stir_coffee/episode_0/cam_high/frame_000015.jpg", "images/human_pika/stir_coffee/episode_0/cam_left_wrist/frame_000015.jpg", "images/human_pika/stir_coffee/episode_0/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stir coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stir coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_16-train-bf_000255-af_000165-minus-0000874", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_16/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_16/cam_high/frame_000562.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_16/cam_high/frame_000255.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_16/cam_left_wrist/frame_000255.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_16/cam_right_wrist/frame_000255.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_16/cam_high/frame_000165.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_16/cam_left_wrist/frame_000165.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_16/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -35.3, "hop_absolute_value": 35.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_peach_to_the_basket_episode_20-train-bf_000195-af_000015-minus-0000009", "task": "put peach to the basket", "image": [ "images/human_pika/put_peach_to_the_basket/episode_22/cam_high/frame_000000.jpg", "images/human_pika/put_peach_to_the_basket/episode_22/cam_high/frame_000498.jpg", "images/human_pika/put_peach_to_the_basket/episode_20/cam_high/frame_000195.jpg", "images/human_pika/put_peach_to_the_basket/episode_20/cam_left_wrist/frame_000195.jpg", "images/human_pika/put_peach_to_the_basket/episode_20/cam_right_wrist/frame_000195.jpg", "images/human_pika/put_peach_to_the_basket/episode_20/cam_high/frame_000015.jpg", "images/human_pika/put_peach_to_the_basket/episode_20/cam_left_wrist/frame_000015.jpg", "images/human_pika/put_peach_to_the_basket/episode_20/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put peach to the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put peach to the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -92.3, "hop_absolute_value": 92.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_plate_storage_episode_17-train-bf_000487-af_000420-minus-0000628", "task": "pour plate storage", "image": [ "images/human_pika/pour_plate_storage/episode_23/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_23/cam_high/frame_000553.jpg", "images/human_pika/pour_plate_storage/episode_17/cam_high/frame_000487.jpg", "images/human_pika/pour_plate_storage/episode_17/cam_left_wrist/frame_000487.jpg", "images/human_pika/pour_plate_storage/episode_17/cam_right_wrist/frame_000487.jpg", "images/human_pika/pour_plate_storage/episode_17/cam_high/frame_000420.jpg", "images/human_pika/pour_plate_storage/episode_17/cam_left_wrist/frame_000420.jpg", "images/human_pika/pour_plate_storage/episode_17/cam_right_wrist/frame_000420.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour plate storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour plate storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_seal_bag_episode_3-train-bf_000210-af_000195-minus-0000609", "task": "seal bag", "image": [ "images/human_pika/seal_bag/episode_3/cam_high/frame_000000.jpg", "images/human_pika/seal_bag/episode_3/cam_high/frame_000535.jpg", "images/human_pika/seal_bag/episode_3/cam_high/frame_000210.jpg", "images/human_pika/seal_bag/episode_3/cam_left_wrist/frame_000210.jpg", "images/human_pika/seal_bag/episode_3/cam_right_wrist/frame_000210.jpg", "images/human_pika/seal_bag/episode_3/cam_high/frame_000195.jpg", "images/human_pika/seal_bag/episode_3/cam_left_wrist/frame_000195.jpg", "images/human_pika/seal_bag/episode_3/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`seal bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `seal bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.1, "hop_absolute_value": 7.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_plate_storage_episode_17-train-bf_000105-af_000060-minus-0000786", "task": "pour plate storage", "image": [ "images/human_pika/pour_plate_storage/episode_4/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_4/cam_high/frame_000481.jpg", "images/human_pika/pour_plate_storage/episode_17/cam_high/frame_000105.jpg", "images/human_pika/pour_plate_storage/episode_17/cam_left_wrist/frame_000105.jpg", "images/human_pika/pour_plate_storage/episode_17/cam_right_wrist/frame_000105.jpg", "images/human_pika/pour_plate_storage/episode_17/cam_high/frame_000060.jpg", "images/human_pika/pour_plate_storage/episode_17/cam_left_wrist/frame_000060.jpg", "images/human_pika/pour_plate_storage/episode_17/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour plate storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour plate storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -42.9, "hop_absolute_value": 42.9, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_plate_storage_episode_8-train-bf_000462-af_000240-minus-0000187", "task": "pour plate storage", "image": [ "images/human_pika/pour_plate_storage/episode_5/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_5/cam_high/frame_000462.jpg", "images/human_pika/pour_plate_storage/episode_8/cam_high/frame_000462.jpg", "images/human_pika/pour_plate_storage/episode_8/cam_left_wrist/frame_000462.jpg", "images/human_pika/pour_plate_storage/episode_8/cam_right_wrist/frame_000462.jpg", "images/human_pika/pour_plate_storage/episode_8/cam_high/frame_000240.jpg", "images/human_pika/pour_plate_storage/episode_8/cam_left_wrist/frame_000240.jpg", "images/human_pika/pour_plate_storage/episode_8/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour plate storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour plate storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -46.7, "hop_absolute_value": 46.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_stack_basket_episode_18-train-bf_000959-af_000105-minus-0000068", "task": "stack basket", "image": [ "images/human_pika/stack_basket/episode_18/cam_high/frame_000000.jpg", "images/human_pika/stack_basket/episode_18/cam_high/frame_000959.jpg", "images/human_pika/stack_basket/episode_18/cam_high/frame_000959.jpg", "images/human_pika/stack_basket/episode_18/cam_left_wrist/frame_000959.jpg", "images/human_pika/stack_basket/episode_18/cam_right_wrist/frame_000959.jpg", "images/human_pika/stack_basket/episode_18/cam_high/frame_000105.jpg", "images/human_pika/stack_basket/episode_18/cam_left_wrist/frame_000105.jpg", "images/human_pika/stack_basket/episode_18/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stack basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stack basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -88.9, "hop_absolute_value": 88.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_fill_basket_with_peach_episode_9-train-bf_000255-af_000045-minus-0000917", "task": "fill basket with peach", "image": [ "images/human_pika/fill_basket_with_peach/episode_21/cam_high/frame_000000.jpg", "images/human_pika/fill_basket_with_peach/episode_21/cam_high/frame_000503.jpg", "images/human_pika/fill_basket_with_peach/episode_9/cam_high/frame_000255.jpg", "images/human_pika/fill_basket_with_peach/episode_9/cam_left_wrist/frame_000255.jpg", "images/human_pika/fill_basket_with_peach/episode_9/cam_right_wrist/frame_000255.jpg", "images/human_pika/fill_basket_with_peach/episode_9/cam_high/frame_000045.jpg", "images/human_pika/fill_basket_with_peach/episode_9/cam_left_wrist/frame_000045.jpg", "images/human_pika/fill_basket_with_peach/episode_9/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fill basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fill basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -82.4, "hop_absolute_value": 82.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_groceries_into_bag_episode_3-train-bf_000390-af_000360-minus-0000607", "task": "put groceries into bag", "image": [ "images/human_pika/put_groceries_into_bag/episode_19/cam_high/frame_000000.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_high/frame_000532.jpg", "images/human_pika/put_groceries_into_bag/episode_3/cam_high/frame_000390.jpg", "images/human_pika/put_groceries_into_bag/episode_3/cam_left_wrist/frame_000390.jpg", "images/human_pika/put_groceries_into_bag/episode_3/cam_right_wrist/frame_000390.jpg", "images/human_pika/put_groceries_into_bag/episode_3/cam_high/frame_000360.jpg", "images/human_pika/put_groceries_into_bag/episode_3/cam_left_wrist/frame_000360.jpg", "images/human_pika/put_groceries_into_bag/episode_3/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put groceries into bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put groceries into bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.7, "hop_absolute_value": 7.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_peach_episode_15-train-bf_000300-af_000240-minus-0000264", "task": "pour basket with peach", "image": [ "images/human_pika/pour_basket_with_peach/episode_8/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_peach/episode_8/cam_high/frame_000523.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_high/frame_000300.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_left_wrist/frame_000300.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_right_wrist/frame_000300.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_high/frame_000240.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_left_wrist/frame_000240.jpg", "images/human_pika/pour_basket_with_peach/episode_15/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_plate_storage_episode_16-train-bf_000240-af_000015-minus-0000017", "task": "pour plate storage", "image": [ "images/human_pika/pour_plate_storage/episode_0/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_0/cam_high/frame_000538.jpg", "images/human_pika/pour_plate_storage/episode_16/cam_high/frame_000240.jpg", "images/human_pika/pour_plate_storage/episode_16/cam_left_wrist/frame_000240.jpg", "images/human_pika/pour_plate_storage/episode_16/cam_right_wrist/frame_000240.jpg", "images/human_pika/pour_plate_storage/episode_16/cam_high/frame_000015.jpg", "images/human_pika/pour_plate_storage/episode_16/cam_left_wrist/frame_000015.jpg", "images/human_pika/pour_plate_storage/episode_16/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour plate storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour plate storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -93.8, "hop_absolute_value": 93.8, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_place_test_tube_episode_21-train-bf_000435-af_000345-minus-0000876", "task": "place test tube", "image": [ "images/human_pika/place_test_tube/episode_13/cam_high/frame_000000.jpg", "images/human_pika/place_test_tube/episode_13/cam_high/frame_000535.jpg", "images/human_pika/place_test_tube/episode_21/cam_high/frame_000435.jpg", "images/human_pika/place_test_tube/episode_21/cam_left_wrist/frame_000435.jpg", "images/human_pika/place_test_tube/episode_21/cam_right_wrist/frame_000435.jpg", "images/human_pika/place_test_tube/episode_21/cam_high/frame_000345.jpg", "images/human_pika/place_test_tube/episode_21/cam_left_wrist/frame_000345.jpg", "images/human_pika/place_test_tube/episode_21/cam_right_wrist/frame_000345.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`place test tube`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `place test tube`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.7, "hop_absolute_value": 20.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_stir_coffee_episode_10-train-bf_000465-af_000210-minus-0000172", "task": "stir coffee", "image": [ "images/human_pika/stir_coffee/episode_7/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_7/cam_high/frame_000524.jpg", "images/human_pika/stir_coffee/episode_10/cam_high/frame_000465.jpg", "images/human_pika/stir_coffee/episode_10/cam_left_wrist/frame_000465.jpg", "images/human_pika/stir_coffee/episode_10/cam_right_wrist/frame_000465.jpg", "images/human_pika/stir_coffee/episode_10/cam_high/frame_000210.jpg", "images/human_pika/stir_coffee/episode_10/cam_left_wrist/frame_000210.jpg", "images/human_pika/stir_coffee/episode_10/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stir coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stir coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -54.8, "hop_absolute_value": 54.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_stir_coffee_episode_18-train-bf_000225-af_000195-minus-0000648", "task": "stir coffee", "image": [ "images/human_pika/stir_coffee/episode_18/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_18/cam_high/frame_000508.jpg", "images/human_pika/stir_coffee/episode_18/cam_high/frame_000225.jpg", "images/human_pika/stir_coffee/episode_18/cam_left_wrist/frame_000225.jpg", "images/human_pika/stir_coffee/episode_18/cam_right_wrist/frame_000225.jpg", "images/human_pika/stir_coffee/episode_18/cam_high/frame_000195.jpg", "images/human_pika/stir_coffee/episode_18/cam_left_wrist/frame_000195.jpg", "images/human_pika/stir_coffee/episode_18/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stir coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stir coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -13.3, "hop_absolute_value": 13.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_16-train-bf_000180-af_000015-minus-0000041", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_22/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_22/cam_high/frame_000654.jpg", "images/human_pika/pour_tea/episode_16/cam_high/frame_000180.jpg", "images/human_pika/pour_tea/episode_16/cam_left_wrist/frame_000180.jpg", "images/human_pika/pour_tea/episode_16/cam_right_wrist/frame_000180.jpg", "images/human_pika/pour_tea/episode_16/cam_high/frame_000015.jpg", "images/human_pika/pour_tea/episode_16/cam_left_wrist/frame_000015.jpg", "images/human_pika/pour_tea/episode_16/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -91.7, "hop_absolute_value": 91.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_wipe_stains_episode_24-train-bf_000435-af_000000-minus-0000019", "task": "wipe stains", "image": [ "images/human_pika/wipe_stains/episode_24/cam_high/frame_000000.jpg", "images/human_pika/wipe_stains/episode_24/cam_high/frame_000608.jpg", "images/human_pika/wipe_stains/episode_24/cam_high/frame_000435.jpg", "images/human_pika/wipe_stains/episode_24/cam_left_wrist/frame_000435.jpg", "images/human_pika/wipe_stains/episode_24/cam_right_wrist/frame_000435.jpg", "images/human_pika/wipe_stains/episode_24/cam_high/frame_000000.jpg", "images/human_pika/wipe_stains/episode_24/cam_left_wrist/frame_000000.jpg", "images/human_pika/wipe_stains/episode_24/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`wipe stains`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `wipe stains`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_stir_coffee_episode_8-train-bf_000525-af_000180-minus-0000155", "task": "stir coffee", "image": [ "images/human_pika/stir_coffee/episode_0/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_0/cam_high/frame_000480.jpg", "images/human_pika/stir_coffee/episode_8/cam_high/frame_000525.jpg", "images/human_pika/stir_coffee/episode_8/cam_left_wrist/frame_000525.jpg", "images/human_pika/stir_coffee/episode_8/cam_right_wrist/frame_000525.jpg", "images/human_pika/stir_coffee/episode_8/cam_high/frame_000180.jpg", "images/human_pika/stir_coffee/episode_8/cam_left_wrist/frame_000180.jpg", "images/human_pika/stir_coffee/episode_8/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stir coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stir coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -65.7, "hop_absolute_value": 65.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_cup_water_containing_episode_16-train-bf_000480-af_000150-minus-0000124", "task": "pour cup water containing", "image": [ "images/human_pika/pour_cup_water_containing/episode_7/cam_high/frame_000000.jpg", "images/human_pika/pour_cup_water_containing/episode_7/cam_high/frame_000492.jpg", "images/human_pika/pour_cup_water_containing/episode_16/cam_high/frame_000480.jpg", "images/human_pika/pour_cup_water_containing/episode_16/cam_left_wrist/frame_000480.jpg", "images/human_pika/pour_cup_water_containing/episode_16/cam_right_wrist/frame_000480.jpg", "images/human_pika/pour_cup_water_containing/episode_16/cam_high/frame_000150.jpg", "images/human_pika/pour_cup_water_containing/episode_16/cam_left_wrist/frame_000150.jpg", "images/human_pika/pour_cup_water_containing/episode_16/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour cup water containing`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour cup water containing`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -68.8, "hop_absolute_value": 68.8, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_place_test_tube_episode_17-train-bf_000540-af_000450-minus-0000731", "task": "place test tube", "image": [ "images/human_pika/place_test_tube/episode_17/cam_high/frame_000000.jpg", "images/human_pika/place_test_tube/episode_17/cam_high/frame_000621.jpg", "images/human_pika/place_test_tube/episode_17/cam_high/frame_000540.jpg", "images/human_pika/place_test_tube/episode_17/cam_left_wrist/frame_000540.jpg", "images/human_pika/place_test_tube/episode_17/cam_right_wrist/frame_000540.jpg", "images/human_pika/place_test_tube/episode_17/cam_high/frame_000450.jpg", "images/human_pika/place_test_tube/episode_17/cam_left_wrist/frame_000450.jpg", "images/human_pika/place_test_tube/episode_17/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`place test tube`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `place test tube`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_scoop_bean_episode_2-train-bf_000525-af_000225-minus-0000191", "task": "scoop bean", "image": [ "images/human_pika/scoop_bean/episode_12/cam_high/frame_000000.jpg", "images/human_pika/scoop_bean/episode_12/cam_high/frame_000530.jpg", "images/human_pika/scoop_bean/episode_2/cam_high/frame_000525.jpg", "images/human_pika/scoop_bean/episode_2/cam_left_wrist/frame_000525.jpg", "images/human_pika/scoop_bean/episode_2/cam_right_wrist/frame_000525.jpg", "images/human_pika/scoop_bean/episode_2/cam_high/frame_000225.jpg", "images/human_pika/scoop_bean/episode_2/cam_left_wrist/frame_000225.jpg", "images/human_pika/scoop_bean/episode_2/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`scoop bean`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `scoop bean`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -57.1, "hop_absolute_value": 57.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_build_the_blocks_episode_10-train-bf_000375-af_000090-minus-0000126", "task": "build the blocks", "image": [ "images/human_pika/build_the_blocks/episode_5/cam_high/frame_000000.jpg", "images/human_pika/build_the_blocks/episode_5/cam_high/frame_000587.jpg", "images/human_pika/build_the_blocks/episode_10/cam_high/frame_000375.jpg", "images/human_pika/build_the_blocks/episode_10/cam_left_wrist/frame_000375.jpg", "images/human_pika/build_the_blocks/episode_10/cam_right_wrist/frame_000375.jpg", "images/human_pika/build_the_blocks/episode_10/cam_high/frame_000090.jpg", "images/human_pika/build_the_blocks/episode_10/cam_left_wrist/frame_000090.jpg", "images/human_pika/build_the_blocks/episode_10/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`build the blocks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `build the blocks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -76.0, "hop_absolute_value": 76.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_test_tube_to_the_shelf_episode_0-train-bf_000240-af_000030-minus-0000047", "task": "put test tube to the shelf", "image": [ "images/human_pika/put_test_tube_to_the_shelf/episode_16/cam_high/frame_000000.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_16/cam_high/frame_000559.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_high/frame_000240.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_left_wrist/frame_000240.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_right_wrist/frame_000240.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_high/frame_000030.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_left_wrist/frame_000030.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_0/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put test tube to the shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put test tube to the shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_build_blocks_episode_24-train-bf_000345-af_000105-minus-0000116", "task": "build blocks", "image": [ "images/human_pika/build_blocks/episode_16/cam_high/frame_000000.jpg", "images/human_pika/build_blocks/episode_16/cam_high/frame_000474.jpg", "images/human_pika/build_blocks/episode_24/cam_high/frame_000345.jpg", "images/human_pika/build_blocks/episode_24/cam_left_wrist/frame_000345.jpg", "images/human_pika/build_blocks/episode_24/cam_right_wrist/frame_000345.jpg", "images/human_pika/build_blocks/episode_24/cam_high/frame_000105.jpg", "images/human_pika/build_blocks/episode_24/cam_left_wrist/frame_000105.jpg", "images/human_pika/build_blocks/episode_24/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`build blocks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `build blocks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -69.6, "hop_absolute_value": 69.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_pants_episode_8-train-bf_000390-af_000060-minus-0000050", "task": "fold pants", "image": [ "images/human_pika/fold_pants/episode_8/cam_high/frame_000000.jpg", "images/human_pika/fold_pants/episode_8/cam_high/frame_000452.jpg", "images/human_pika/fold_pants/episode_8/cam_high/frame_000390.jpg", "images/human_pika/fold_pants/episode_8/cam_left_wrist/frame_000390.jpg", "images/human_pika/fold_pants/episode_8/cam_right_wrist/frame_000390.jpg", "images/human_pika/fold_pants/episode_8/cam_high/frame_000060.jpg", "images/human_pika/fold_pants/episode_8/cam_left_wrist/frame_000060.jpg", "images/human_pika/fold_pants/episode_8/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold pants`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold pants`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -84.6, "hop_absolute_value": 84.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_23-train-bf_000225-af_000210-minus-0000548", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_16/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_16/cam_high/frame_000562.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_high/frame_000225.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_left_wrist/frame_000225.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_right_wrist/frame_000225.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_high/frame_000210.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_left_wrist/frame_000210.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.7, "hop_absolute_value": 6.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_build_the_blocks_episode_5-train-bf_000587-af_000570-minus-0000584", "task": "build the blocks", "image": [ "images/human_pika/build_the_blocks/episode_16/cam_high/frame_000000.jpg", "images/human_pika/build_the_blocks/episode_16/cam_high/frame_000547.jpg", "images/human_pika/build_the_blocks/episode_5/cam_high/frame_000587.jpg", "images/human_pika/build_the_blocks/episode_5/cam_left_wrist/frame_000587.jpg", "images/human_pika/build_the_blocks/episode_5/cam_right_wrist/frame_000587.jpg", "images/human_pika/build_the_blocks/episode_5/cam_high/frame_000570.jpg", "images/human_pika/build_the_blocks/episode_5/cam_left_wrist/frame_000570.jpg", "images/human_pika/build_the_blocks/episode_5/cam_right_wrist/frame_000570.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`build the blocks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `build the blocks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.6, "hop_absolute_value": 2.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_4-train-bf_000285-af_000030-minus-0000052", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_4/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_4/cam_high/frame_000663.jpg", "images/human_pika/pour_tea/episode_4/cam_high/frame_000285.jpg", "images/human_pika/pour_tea/episode_4/cam_left_wrist/frame_000285.jpg", "images/human_pika/pour_tea/episode_4/cam_right_wrist/frame_000285.jpg", "images/human_pika/pour_tea/episode_4/cam_high/frame_000030.jpg", "images/human_pika/pour_tea/episode_4/cam_left_wrist/frame_000030.jpg", "images/human_pika/pour_tea/episode_4/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -89.5, "hop_absolute_value": 89.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_build_the_blocks_episode_3-train-bf_000255-af_000045-minus-0000081", "task": "build the blocks", "image": [ "images/human_pika/build_the_blocks/episode_9/cam_high/frame_000000.jpg", "images/human_pika/build_the_blocks/episode_9/cam_high/frame_000616.jpg", "images/human_pika/build_the_blocks/episode_3/cam_high/frame_000255.jpg", "images/human_pika/build_the_blocks/episode_3/cam_left_wrist/frame_000255.jpg", "images/human_pika/build_the_blocks/episode_3/cam_right_wrist/frame_000255.jpg", "images/human_pika/build_the_blocks/episode_3/cam_high/frame_000045.jpg", "images/human_pika/build_the_blocks/episode_3/cam_left_wrist/frame_000045.jpg", "images/human_pika/build_the_blocks/episode_3/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`build the blocks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `build the blocks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -82.4, "hop_absolute_value": 82.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bread_to_basket_episode_15-train-bf_000405-af_000390-minus-0000586", "task": "pour bread to basket", "image": [ "images/human_pika/pour_bread_to_basket/episode_1/cam_high/frame_000000.jpg", "images/human_pika/pour_bread_to_basket/episode_1/cam_high/frame_000521.jpg", "images/human_pika/pour_bread_to_basket/episode_15/cam_high/frame_000405.jpg", "images/human_pika/pour_bread_to_basket/episode_15/cam_left_wrist/frame_000405.jpg", "images/human_pika/pour_bread_to_basket/episode_15/cam_right_wrist/frame_000405.jpg", "images/human_pika/pour_bread_to_basket/episode_15/cam_high/frame_000390.jpg", "images/human_pika/pour_bread_to_basket/episode_15/cam_left_wrist/frame_000390.jpg", "images/human_pika/pour_bread_to_basket/episode_15/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bread to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bread to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.7, "hop_absolute_value": 3.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_build_the_blocks_episode_2-train-bf_000285-af_000270-minus-0000662", "task": "build the blocks", "image": [ "images/human_pika/build_the_blocks/episode_12/cam_high/frame_000000.jpg", "images/human_pika/build_the_blocks/episode_12/cam_high/frame_000655.jpg", "images/human_pika/build_the_blocks/episode_2/cam_high/frame_000285.jpg", "images/human_pika/build_the_blocks/episode_2/cam_left_wrist/frame_000285.jpg", "images/human_pika/build_the_blocks/episode_2/cam_right_wrist/frame_000285.jpg", "images/human_pika/build_the_blocks/episode_2/cam_high/frame_000270.jpg", "images/human_pika/build_the_blocks/episode_2/cam_left_wrist/frame_000270.jpg", "images/human_pika/build_the_blocks/episode_2/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`build the blocks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `build the blocks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.3, "hop_absolute_value": 5.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_test_tube_to_the_shelf_episode_17-train-bf_000621-af_000270-minus-0000188", "task": "put test tube to the shelf", "image": [ "images/human_pika/put_test_tube_to_the_shelf/episode_20/cam_high/frame_000000.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_20/cam_high/frame_000677.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_high/frame_000621.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_left_wrist/frame_000621.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_right_wrist/frame_000621.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_high/frame_000270.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_left_wrist/frame_000270.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_17/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put test tube to the shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put test tube to the shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -56.1, "hop_absolute_value": 56.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_peach_to_the_basket_episode_3-train-bf_000090-af_000015-minus-0000073", "task": "put peach to the basket", "image": [ "images/human_pika/put_peach_to_the_basket/episode_17/cam_high/frame_000000.jpg", "images/human_pika/put_peach_to_the_basket/episode_17/cam_high/frame_000480.jpg", "images/human_pika/put_peach_to_the_basket/episode_3/cam_high/frame_000090.jpg", "images/human_pika/put_peach_to_the_basket/episode_3/cam_left_wrist/frame_000090.jpg", "images/human_pika/put_peach_to_the_basket/episode_3/cam_right_wrist/frame_000090.jpg", "images/human_pika/put_peach_to_the_basket/episode_3/cam_high/frame_000015.jpg", "images/human_pika/put_peach_to_the_basket/episode_3/cam_left_wrist/frame_000015.jpg", "images/human_pika/put_peach_to_the_basket/episode_3/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put peach to the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put peach to the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_23-train-bf_000525-af_000360-minus-0000247", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_1/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_1/cam_high/frame_000495.jpg", "images/human_pika/pour_tea/episode_23/cam_high/frame_000525.jpg", "images/human_pika/pour_tea/episode_23/cam_left_wrist/frame_000525.jpg", "images/human_pika/pour_tea/episode_23/cam_right_wrist/frame_000525.jpg", "images/human_pika/pour_tea/episode_23/cam_high/frame_000360.jpg", "images/human_pika/pour_tea/episode_23/cam_left_wrist/frame_000360.jpg", "images/human_pika/pour_tea/episode_23/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -31.4, "hop_absolute_value": 31.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_build_blocks_episode_7-train-bf_000330-af_000225-minus-0000225", "task": "build blocks", "image": [ "images/human_pika/build_blocks/episode_7/cam_high/frame_000000.jpg", "images/human_pika/build_blocks/episode_7/cam_high/frame_000498.jpg", "images/human_pika/build_blocks/episode_7/cam_high/frame_000330.jpg", "images/human_pika/build_blocks/episode_7/cam_left_wrist/frame_000330.jpg", "images/human_pika/build_blocks/episode_7/cam_right_wrist/frame_000330.jpg", "images/human_pika/build_blocks/episode_7/cam_high/frame_000225.jpg", "images/human_pika/build_blocks/episode_7/cam_left_wrist/frame_000225.jpg", "images/human_pika/build_blocks/episode_7/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`build blocks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `build blocks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -31.8, "hop_absolute_value": 31.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_peach_episode_20-train-bf_000480-af_000450-minus-0000610", "task": "pour basket with peach", "image": [ "images/human_pika/pour_basket_with_peach/episode_24/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_peach/episode_24/cam_high/frame_000542.jpg", "images/human_pika/pour_basket_with_peach/episode_20/cam_high/frame_000480.jpg", "images/human_pika/pour_basket_with_peach/episode_20/cam_left_wrist/frame_000480.jpg", "images/human_pika/pour_basket_with_peach/episode_20/cam_right_wrist/frame_000480.jpg", "images/human_pika/pour_basket_with_peach/episode_20/cam_high/frame_000450.jpg", "images/human_pika/pour_basket_with_peach/episode_20/cam_left_wrist/frame_000450.jpg", "images/human_pika/pour_basket_with_peach/episode_20/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.2, "hop_absolute_value": 6.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_plate_storage_episode_3-train-bf_000525-af_000360-minus-0000252", "task": "pour plate storage", "image": [ "images/human_pika/pour_plate_storage/episode_4/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_4/cam_high/frame_000481.jpg", "images/human_pika/pour_plate_storage/episode_3/cam_high/frame_000525.jpg", "images/human_pika/pour_plate_storage/episode_3/cam_left_wrist/frame_000525.jpg", "images/human_pika/pour_plate_storage/episode_3/cam_right_wrist/frame_000525.jpg", "images/human_pika/pour_plate_storage/episode_3/cam_high/frame_000360.jpg", "images/human_pika/pour_plate_storage/episode_3/cam_left_wrist/frame_000360.jpg", "images/human_pika/pour_plate_storage/episode_3/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour plate storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour plate storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -31.4, "hop_absolute_value": 31.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_stir_coffee_episode_10-train-bf_000420-af_000405-minus-0000581", "task": "stir coffee", "image": [ "images/human_pika/stir_coffee/episode_2/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_2/cam_high/frame_000567.jpg", "images/human_pika/stir_coffee/episode_10/cam_high/frame_000420.jpg", "images/human_pika/stir_coffee/episode_10/cam_left_wrist/frame_000420.jpg", "images/human_pika/stir_coffee/episode_10/cam_right_wrist/frame_000420.jpg", "images/human_pika/stir_coffee/episode_10/cam_high/frame_000405.jpg", "images/human_pika/stir_coffee/episode_10/cam_left_wrist/frame_000405.jpg", "images/human_pika/stir_coffee/episode_10/cam_right_wrist/frame_000405.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stir coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stir coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.6, "hop_absolute_value": 3.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_cup_water_containing_episode_6-train-bf_000330-af_000285-minus-0000596", "task": "pour cup water containing", "image": [ "images/human_pika/pour_cup_water_containing/episode_0/cam_high/frame_000000.jpg", "images/human_pika/pour_cup_water_containing/episode_0/cam_high/frame_000600.jpg", "images/human_pika/pour_cup_water_containing/episode_6/cam_high/frame_000330.jpg", "images/human_pika/pour_cup_water_containing/episode_6/cam_left_wrist/frame_000330.jpg", "images/human_pika/pour_cup_water_containing/episode_6/cam_right_wrist/frame_000330.jpg", "images/human_pika/pour_cup_water_containing/episode_6/cam_high/frame_000285.jpg", "images/human_pika/pour_cup_water_containing/episode_6/cam_left_wrist/frame_000285.jpg", "images/human_pika/pour_cup_water_containing/episode_6/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour cup water containing`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour cup water containing`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -13.6, "hop_absolute_value": 13.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_scoop_bean_episode_17-train-bf_000225-af_000135-minus-0000219", "task": "scoop bean", "image": [ "images/human_pika/scoop_bean/episode_15/cam_high/frame_000000.jpg", "images/human_pika/scoop_bean/episode_15/cam_high/frame_000494.jpg", "images/human_pika/scoop_bean/episode_17/cam_high/frame_000225.jpg", "images/human_pika/scoop_bean/episode_17/cam_left_wrist/frame_000225.jpg", "images/human_pika/scoop_bean/episode_17/cam_right_wrist/frame_000225.jpg", "images/human_pika/scoop_bean/episode_17/cam_high/frame_000135.jpg", "images/human_pika/scoop_bean/episode_17/cam_left_wrist/frame_000135.jpg", "images/human_pika/scoop_bean/episode_17/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`scoop bean`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `scoop bean`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bowl_episode_3-train-bf_000255-af_000225-minus-0000762", "task": "pour bowl", "image": [ "images/human_pika/pour_bowl/episode_6/cam_high/frame_000000.jpg", "images/human_pika/pour_bowl/episode_6/cam_high/frame_000906.jpg", "images/human_pika/pour_bowl/episode_3/cam_high/frame_000255.jpg", "images/human_pika/pour_bowl/episode_3/cam_left_wrist/frame_000255.jpg", "images/human_pika/pour_bowl/episode_3/cam_right_wrist/frame_000255.jpg", "images/human_pika/pour_bowl/episode_3/cam_high/frame_000225.jpg", "images/human_pika/pour_bowl/episode_3/cam_left_wrist/frame_000225.jpg", "images/human_pika/pour_bowl/episode_3/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.8, "hop_absolute_value": 11.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_plate_storage_episode_24-train-bf_000180-af_000000-minus-0000009", "task": "pour plate storage", "image": [ "images/human_pika/pour_plate_storage/episode_11/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_11/cam_high/frame_000514.jpg", "images/human_pika/pour_plate_storage/episode_24/cam_high/frame_000180.jpg", "images/human_pika/pour_plate_storage/episode_24/cam_left_wrist/frame_000180.jpg", "images/human_pika/pour_plate_storage/episode_24/cam_right_wrist/frame_000180.jpg", "images/human_pika/pour_plate_storage/episode_24/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_24/cam_left_wrist/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_24/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour plate storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour plate storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_seal_bag_episode_24-train-bf_000540-af_000510-minus-0000619", "task": "seal bag", "image": [ "images/human_pika/seal_bag/episode_7/cam_high/frame_000000.jpg", "images/human_pika/seal_bag/episode_7/cam_high/frame_000537.jpg", "images/human_pika/seal_bag/episode_24/cam_high/frame_000540.jpg", "images/human_pika/seal_bag/episode_24/cam_left_wrist/frame_000540.jpg", "images/human_pika/seal_bag/episode_24/cam_right_wrist/frame_000540.jpg", "images/human_pika/seal_bag/episode_24/cam_high/frame_000510.jpg", "images/human_pika/seal_bag/episode_24/cam_left_wrist/frame_000510.jpg", "images/human_pika/seal_bag/episode_24/cam_right_wrist/frame_000510.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`seal bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `seal bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.6, "hop_absolute_value": 5.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_13-train-bf_000045-af_000015-minus-0000133", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_19/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_19/cam_high/frame_000529.jpg", "images/human_pika/pour_tea/episode_13/cam_high/frame_000045.jpg", "images/human_pika/pour_tea/episode_13/cam_left_wrist/frame_000045.jpg", "images/human_pika/pour_tea/episode_13/cam_right_wrist/frame_000045.jpg", "images/human_pika/pour_tea/episode_13/cam_high/frame_000015.jpg", "images/human_pika/pour_tea/episode_13/cam_left_wrist/frame_000015.jpg", "images/human_pika/pour_tea/episode_13/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_17-train-bf_000285-af_000030-minus-0000048", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_1/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_1/cam_high/frame_000575.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_high/frame_000285.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_left_wrist/frame_000285.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_right_wrist/frame_000285.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_high/frame_000030.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_left_wrist/frame_000030.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -89.5, "hop_absolute_value": 89.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_17-train-bf_000465-af_000135-minus-0000124", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_3/cam_high/frame_000543.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_high/frame_000465.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_left_wrist/frame_000465.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_right_wrist/frame_000465.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_high/frame_000135.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_left_wrist/frame_000135.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_17/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -71.0, "hop_absolute_value": 71.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_scoop_bean_episode_18-train-bf_000285-af_000045-minus-0000050", "task": "scoop bean", "image": [ "images/human_pika/scoop_bean/episode_18/cam_high/frame_000000.jpg", "images/human_pika/scoop_bean/episode_18/cam_high/frame_000552.jpg", "images/human_pika/scoop_bean/episode_18/cam_high/frame_000285.jpg", "images/human_pika/scoop_bean/episode_18/cam_left_wrist/frame_000285.jpg", "images/human_pika/scoop_bean/episode_18/cam_right_wrist/frame_000285.jpg", "images/human_pika/scoop_bean/episode_18/cam_high/frame_000045.jpg", "images/human_pika/scoop_bean/episode_18/cam_left_wrist/frame_000045.jpg", "images/human_pika/scoop_bean/episode_18/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`scoop bean`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `scoop bean`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -84.2, "hop_absolute_value": 84.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_stir_coffee_episode_20-train-bf_000435-af_000315-minus-0000855", "task": "stir coffee", "image": [ "images/human_pika/stir_coffee/episode_5/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_5/cam_high/frame_000575.jpg", "images/human_pika/stir_coffee/episode_20/cam_high/frame_000435.jpg", "images/human_pika/stir_coffee/episode_20/cam_left_wrist/frame_000435.jpg", "images/human_pika/stir_coffee/episode_20/cam_right_wrist/frame_000435.jpg", "images/human_pika/stir_coffee/episode_20/cam_high/frame_000315.jpg", "images/human_pika/stir_coffee/episode_20/cam_left_wrist/frame_000315.jpg", "images/human_pika/stir_coffee/episode_20/cam_right_wrist/frame_000315.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stir coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stir coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -27.6, "hop_absolute_value": 27.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_stack_basket_episode_16-train-bf_000195-af_000090-minus-0000178", "task": "stack basket", "image": [ "images/human_pika/stack_basket/episode_16/cam_high/frame_000000.jpg", "images/human_pika/stack_basket/episode_16/cam_high/frame_000978.jpg", "images/human_pika/stack_basket/episode_16/cam_high/frame_000195.jpg", "images/human_pika/stack_basket/episode_16/cam_left_wrist/frame_000195.jpg", "images/human_pika/stack_basket/episode_16/cam_right_wrist/frame_000195.jpg", "images/human_pika/stack_basket/episode_16/cam_high/frame_000090.jpg", "images/human_pika/stack_basket/episode_16/cam_left_wrist/frame_000090.jpg", "images/human_pika/stack_basket/episode_16/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stack basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stack basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -53.8, "hop_absolute_value": 53.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_cup_water_containing_episode_4-train-bf_000360-af_000210-minus-0000220", "task": "pour cup water containing", "image": [ "images/human_pika/pour_cup_water_containing/episode_8/cam_high/frame_000000.jpg", "images/human_pika/pour_cup_water_containing/episode_8/cam_high/frame_000469.jpg", "images/human_pika/pour_cup_water_containing/episode_4/cam_high/frame_000360.jpg", "images/human_pika/pour_cup_water_containing/episode_4/cam_left_wrist/frame_000360.jpg", "images/human_pika/pour_cup_water_containing/episode_4/cam_right_wrist/frame_000360.jpg", "images/human_pika/pour_cup_water_containing/episode_4/cam_high/frame_000210.jpg", "images/human_pika/pour_cup_water_containing/episode_4/cam_left_wrist/frame_000210.jpg", "images/human_pika/pour_cup_water_containing/episode_4/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour cup water containing`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour cup water containing`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -41.7, "hop_absolute_value": 41.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_seal_bag_episode_19-train-bf_000390-af_000270-minus-0000219", "task": "seal bag", "image": [ "images/human_pika/seal_bag/episode_14/cam_high/frame_000000.jpg", "images/human_pika/seal_bag/episode_14/cam_high/frame_000542.jpg", "images/human_pika/seal_bag/episode_19/cam_high/frame_000390.jpg", "images/human_pika/seal_bag/episode_19/cam_left_wrist/frame_000390.jpg", "images/human_pika/seal_bag/episode_19/cam_right_wrist/frame_000390.jpg", "images/human_pika/seal_bag/episode_19/cam_high/frame_000270.jpg", "images/human_pika/seal_bag/episode_19/cam_left_wrist/frame_000270.jpg", "images/human_pika/seal_bag/episode_19/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`seal bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `seal bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -30.8, "hop_absolute_value": 30.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_groceries_into_bag_episode_19-train-bf_000405-af_000345-minus-0000666", "task": "put groceries into bag", "image": [ "images/human_pika/put_groceries_into_bag/episode_19/cam_high/frame_000000.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_high/frame_000532.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_high/frame_000405.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_left_wrist/frame_000405.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_right_wrist/frame_000405.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_high/frame_000345.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_left_wrist/frame_000345.jpg", "images/human_pika/put_groceries_into_bag/episode_19/cam_right_wrist/frame_000345.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put groceries into bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put groceries into bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -14.8, "hop_absolute_value": 14.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_towel_episode_9-train-bf_000525-af_000450-minus-0000713", "task": "fold towel", "image": [ "images/human_pika/fold_towel/episode_1/cam_high/frame_000000.jpg", "images/human_pika/fold_towel/episode_1/cam_high/frame_000574.jpg", "images/human_pika/fold_towel/episode_9/cam_high/frame_000525.jpg", "images/human_pika/fold_towel/episode_9/cam_left_wrist/frame_000525.jpg", "images/human_pika/fold_towel/episode_9/cam_right_wrist/frame_000525.jpg", "images/human_pika/fold_towel/episode_9/cam_high/frame_000450.jpg", "images/human_pika/fold_towel/episode_9/cam_left_wrist/frame_000450.jpg", "images/human_pika/fold_towel/episode_9/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold towel`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold towel`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -14.3, "hop_absolute_value": 14.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_test_tube_to_the_shelf_episode_10-train-bf_000390-af_000015-minus-0000850", "task": "put test tube to the shelf", "image": [ "images/human_pika/put_test_tube_to_the_shelf/episode_8/cam_high/frame_000000.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_8/cam_high/frame_000500.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_10/cam_high/frame_000390.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_10/cam_left_wrist/frame_000390.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_10/cam_right_wrist/frame_000390.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_10/cam_high/frame_000015.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_10/cam_left_wrist/frame_000015.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_10/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put test tube to the shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put test tube to the shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -96.2, "hop_absolute_value": 96.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_plate_storage_episode_1-train-bf_000300-af_000285-minus-0000270", "task": "pour plate storage", "image": [ "images/human_pika/pour_plate_storage/episode_3/cam_high/frame_000000.jpg", "images/human_pika/pour_plate_storage/episode_3/cam_high/frame_000594.jpg", "images/human_pika/pour_plate_storage/episode_1/cam_high/frame_000300.jpg", "images/human_pika/pour_plate_storage/episode_1/cam_left_wrist/frame_000300.jpg", "images/human_pika/pour_plate_storage/episode_1/cam_right_wrist/frame_000300.jpg", "images/human_pika/pour_plate_storage/episode_1/cam_high/frame_000285.jpg", "images/human_pika/pour_plate_storage/episode_1/cam_left_wrist/frame_000285.jpg", "images/human_pika/pour_plate_storage/episode_1/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour plate storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour plate storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.0, "hop_absolute_value": 5.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_23-train-bf_000405-af_000330-minus-0000747", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_23/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_23/cam_high/frame_000656.jpg", "images/human_pika/pour_tea/episode_23/cam_high/frame_000405.jpg", "images/human_pika/pour_tea/episode_23/cam_left_wrist/frame_000405.jpg", "images/human_pika/pour_tea/episode_23/cam_right_wrist/frame_000405.jpg", "images/human_pika/pour_tea/episode_23/cam_high/frame_000330.jpg", "images/human_pika/pour_tea/episode_23/cam_left_wrist/frame_000330.jpg", "images/human_pika/pour_tea/episode_23/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -18.5, "hop_absolute_value": 18.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_23-train-bf_000270-af_000195-minus-0000908", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_23/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_23/cam_high/frame_000656.jpg", "images/human_pika/pour_tea/episode_23/cam_high/frame_000270.jpg", "images/human_pika/pour_tea/episode_23/cam_left_wrist/frame_000270.jpg", "images/human_pika/pour_tea/episode_23/cam_right_wrist/frame_000270.jpg", "images/human_pika/pour_tea/episode_23/cam_high/frame_000195.jpg", "images/human_pika/pour_tea/episode_23/cam_left_wrist/frame_000195.jpg", "images/human_pika/pour_tea/episode_23/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -27.8, "hop_absolute_value": 27.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_cup_water_containing_episode_9-train-bf_000420-af_000330-minus-0000646", "task": "pour cup water containing", "image": [ "images/human_pika/pour_cup_water_containing/episode_2/cam_high/frame_000000.jpg", "images/human_pika/pour_cup_water_containing/episode_2/cam_high/frame_000494.jpg", "images/human_pika/pour_cup_water_containing/episode_9/cam_high/frame_000420.jpg", "images/human_pika/pour_cup_water_containing/episode_9/cam_left_wrist/frame_000420.jpg", "images/human_pika/pour_cup_water_containing/episode_9/cam_right_wrist/frame_000420.jpg", "images/human_pika/pour_cup_water_containing/episode_9/cam_high/frame_000330.jpg", "images/human_pika/pour_cup_water_containing/episode_9/cam_left_wrist/frame_000330.jpg", "images/human_pika/pour_cup_water_containing/episode_9/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour cup water containing`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour cup water containing`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -21.4, "hop_absolute_value": 21.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_cup_water_containing_episode_6-train-bf_000345-af_000255-minus-0000241", "task": "pour cup water containing", "image": [ "images/human_pika/pour_cup_water_containing/episode_3/cam_high/frame_000000.jpg", "images/human_pika/pour_cup_water_containing/episode_3/cam_high/frame_000535.jpg", "images/human_pika/pour_cup_water_containing/episode_6/cam_high/frame_000345.jpg", "images/human_pika/pour_cup_water_containing/episode_6/cam_left_wrist/frame_000345.jpg", "images/human_pika/pour_cup_water_containing/episode_6/cam_right_wrist/frame_000345.jpg", "images/human_pika/pour_cup_water_containing/episode_6/cam_high/frame_000255.jpg", "images/human_pika/pour_cup_water_containing/episode_6/cam_left_wrist/frame_000255.jpg", "images/human_pika/pour_cup_water_containing/episode_6/cam_right_wrist/frame_000255.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour cup water containing`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour cup water containing`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -26.1, "hop_absolute_value": 26.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_peach_to_the_basket_episode_24-train-bf_000225-af_000180-minus-0000696", "task": "put peach to the basket", "image": [ "images/human_pika/put_peach_to_the_basket/episode_11/cam_high/frame_000000.jpg", "images/human_pika/put_peach_to_the_basket/episode_11/cam_high/frame_000535.jpg", "images/human_pika/put_peach_to_the_basket/episode_24/cam_high/frame_000225.jpg", "images/human_pika/put_peach_to_the_basket/episode_24/cam_left_wrist/frame_000225.jpg", "images/human_pika/put_peach_to_the_basket/episode_24/cam_right_wrist/frame_000225.jpg", "images/human_pika/put_peach_to_the_basket/episode_24/cam_high/frame_000180.jpg", "images/human_pika/put_peach_to_the_basket/episode_24/cam_left_wrist/frame_000180.jpg", "images/human_pika/put_peach_to_the_basket/episode_24/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put peach to the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put peach to the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_stack_basket_episode_1-train-bf_000330-af_000105-minus-0000115", "task": "stack basket", "image": [ "images/human_pika/stack_basket/episode_1/cam_high/frame_000000.jpg", "images/human_pika/stack_basket/episode_1/cam_high/frame_000888.jpg", "images/human_pika/stack_basket/episode_1/cam_high/frame_000330.jpg", "images/human_pika/stack_basket/episode_1/cam_left_wrist/frame_000330.jpg", "images/human_pika/stack_basket/episode_1/cam_right_wrist/frame_000330.jpg", "images/human_pika/stack_basket/episode_1/cam_high/frame_000105.jpg", "images/human_pika/stack_basket/episode_1/cam_left_wrist/frame_000105.jpg", "images/human_pika/stack_basket/episode_1/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stack basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stack basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -68.2, "hop_absolute_value": 68.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_seal_bag_episode_21-train-bf_000542-af_000480-minus-0000582", "task": "seal bag", "image": [ "images/human_pika/seal_bag/episode_20/cam_high/frame_000000.jpg", "images/human_pika/seal_bag/episode_20/cam_high/frame_000459.jpg", "images/human_pika/seal_bag/episode_21/cam_high/frame_000542.jpg", "images/human_pika/seal_bag/episode_21/cam_left_wrist/frame_000542.jpg", "images/human_pika/seal_bag/episode_21/cam_right_wrist/frame_000542.jpg", "images/human_pika/seal_bag/episode_21/cam_high/frame_000480.jpg", "images/human_pika/seal_bag/episode_21/cam_left_wrist/frame_000480.jpg", "images/human_pika/seal_bag/episode_21/cam_right_wrist/frame_000480.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`seal bag`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `seal bag`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bowl_episode_9-train-bf_000840-af_000075-minus-0000069", "task": "pour bowl", "image": [ "images/human_pika/pour_bowl/episode_12/cam_high/frame_000000.jpg", "images/human_pika/pour_bowl/episode_12/cam_high/frame_000877.jpg", "images/human_pika/pour_bowl/episode_9/cam_high/frame_000840.jpg", "images/human_pika/pour_bowl/episode_9/cam_left_wrist/frame_000840.jpg", "images/human_pika/pour_bowl/episode_9/cam_right_wrist/frame_000840.jpg", "images/human_pika/pour_bowl/episode_9/cam_high/frame_000075.jpg", "images/human_pika/pour_bowl/episode_9/cam_left_wrist/frame_000075.jpg", "images/human_pika/pour_bowl/episode_9/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -91.1, "hop_absolute_value": 91.1, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_11-train-bf_000525-af_000330-minus-0000924", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_23/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_23/cam_high/frame_000656.jpg", "images/human_pika/pour_tea/episode_11/cam_high/frame_000525.jpg", "images/human_pika/pour_tea/episode_11/cam_left_wrist/frame_000525.jpg", "images/human_pika/pour_tea/episode_11/cam_right_wrist/frame_000525.jpg", "images/human_pika/pour_tea/episode_11/cam_high/frame_000330.jpg", "images/human_pika/pour_tea/episode_11/cam_left_wrist/frame_000330.jpg", "images/human_pika/pour_tea/episode_11/cam_right_wrist/frame_000330.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -37.1, "hop_absolute_value": 37.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_peach_episode_17-train-bf_000360-af_000285-minus-0000242", "task": "pour basket with peach", "image": [ "images/human_pika/pour_basket_with_peach/episode_2/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_peach/episode_2/cam_high/frame_000541.jpg", "images/human_pika/pour_basket_with_peach/episode_17/cam_high/frame_000360.jpg", "images/human_pika/pour_basket_with_peach/episode_17/cam_left_wrist/frame_000360.jpg", "images/human_pika/pour_basket_with_peach/episode_17/cam_right_wrist/frame_000360.jpg", "images/human_pika/pour_basket_with_peach/episode_17/cam_high/frame_000285.jpg", "images/human_pika/pour_basket_with_peach/episode_17/cam_left_wrist/frame_000285.jpg", "images/human_pika/pour_basket_with_peach/episode_17/cam_right_wrist/frame_000285.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.8, "hop_absolute_value": 20.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_23-train-bf_000270-af_000165-minus-0000797", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_20/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_20/cam_high/frame_000528.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_high/frame_000270.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_left_wrist/frame_000270.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_right_wrist/frame_000270.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_high/frame_000165.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_left_wrist/frame_000165.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -38.9, "hop_absolute_value": 38.9, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_stir_coffee_episode_8-train-bf_000315-af_000300-minus-0000599", "task": "stir coffee", "image": [ "images/human_pika/stir_coffee/episode_8/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_8/cam_high/frame_000610.jpg", "images/human_pika/stir_coffee/episode_8/cam_high/frame_000315.jpg", "images/human_pika/stir_coffee/episode_8/cam_left_wrist/frame_000315.jpg", "images/human_pika/stir_coffee/episode_8/cam_right_wrist/frame_000315.jpg", "images/human_pika/stir_coffee/episode_8/cam_high/frame_000300.jpg", "images/human_pika/stir_coffee/episode_8/cam_left_wrist/frame_000300.jpg", "images/human_pika/stir_coffee/episode_8/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stir coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stir coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.8, "hop_absolute_value": 4.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_egg_yolk_pastry_to_basket_episode_23-train-bf_000090-af_000060-minus-0000217", "task": "pour egg yolk pastry to basket", "image": [ "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_2/cam_high/frame_000000.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_2/cam_high/frame_000499.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_high/frame_000090.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_left_wrist/frame_000090.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_right_wrist/frame_000090.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_high/frame_000060.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_left_wrist/frame_000060.jpg", "images/human_pika/pour_egg_yolk_pastry_to_basket/episode_23/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour egg yolk pastry to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour egg yolk pastry to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_test_tube_to_the_shelf_episode_22-train-bf_000420-af_000300-minus-0000727", "task": "put test tube to the shelf", "image": [ "images/human_pika/put_test_tube_to_the_shelf/episode_1/cam_high/frame_000000.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_1/cam_high/frame_000591.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_22/cam_high/frame_000420.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_22/cam_left_wrist/frame_000420.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_22/cam_right_wrist/frame_000420.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_22/cam_high/frame_000300.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_22/cam_left_wrist/frame_000300.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_22/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put test tube to the shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put test tube to the shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -28.6, "hop_absolute_value": 28.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_wipe_stains_episode_16-train-bf_000390-af_000375-minus-0000284", "task": "wipe stains", "image": [ "images/human_pika/wipe_stains/episode_17/cam_high/frame_000000.jpg", "images/human_pika/wipe_stains/episode_17/cam_high/frame_000456.jpg", "images/human_pika/wipe_stains/episode_16/cam_high/frame_000390.jpg", "images/human_pika/wipe_stains/episode_16/cam_left_wrist/frame_000390.jpg", "images/human_pika/wipe_stains/episode_16/cam_right_wrist/frame_000390.jpg", "images/human_pika/wipe_stains/episode_16/cam_high/frame_000375.jpg", "images/human_pika/wipe_stains/episode_16/cam_left_wrist/frame_000375.jpg", "images/human_pika/wipe_stains/episode_16/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`wipe stains`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `wipe stains`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.8, "hop_absolute_value": 3.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_build_blocks_episode_24-train-bf_000270-af_000045-minus-0000084", "task": "build blocks", "image": [ "images/human_pika/build_blocks/episode_0/cam_high/frame_000000.jpg", "images/human_pika/build_blocks/episode_0/cam_high/frame_000523.jpg", "images/human_pika/build_blocks/episode_24/cam_high/frame_000270.jpg", "images/human_pika/build_blocks/episode_24/cam_left_wrist/frame_000270.jpg", "images/human_pika/build_blocks/episode_24/cam_right_wrist/frame_000270.jpg", "images/human_pika/build_blocks/episode_24/cam_high/frame_000045.jpg", "images/human_pika/build_blocks/episode_24/cam_left_wrist/frame_000045.jpg", "images/human_pika/build_blocks/episode_24/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`build blocks`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `build blocks`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_banana_episode_22-train-bf_000480-af_000390-minus-0000688", "task": "pour basket with banana", "image": [ "images/human_pika/pour_basket_with_banana/episode_24/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_banana/episode_24/cam_high/frame_000513.jpg", "images/human_pika/pour_basket_with_banana/episode_22/cam_high/frame_000480.jpg", "images/human_pika/pour_basket_with_banana/episode_22/cam_left_wrist/frame_000480.jpg", "images/human_pika/pour_basket_with_banana/episode_22/cam_right_wrist/frame_000480.jpg", "images/human_pika/pour_basket_with_banana/episode_22/cam_high/frame_000390.jpg", "images/human_pika/pour_basket_with_banana/episode_22/cam_left_wrist/frame_000390.jpg", "images/human_pika/pour_basket_with_banana/episode_22/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with banana`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with banana`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -18.8, "hop_absolute_value": 18.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_orange_storage_episode_13-train-bf_000255-af_000015-minus-0000014", "task": "orange storage", "image": [ "images/human_pika/orange_storage/episode_11/cam_high/frame_000000.jpg", "images/human_pika/orange_storage/episode_11/cam_high/frame_000572.jpg", "images/human_pika/orange_storage/episode_13/cam_high/frame_000255.jpg", "images/human_pika/orange_storage/episode_13/cam_left_wrist/frame_000255.jpg", "images/human_pika/orange_storage/episode_13/cam_right_wrist/frame_000255.jpg", "images/human_pika/orange_storage/episode_13/cam_high/frame_000015.jpg", "images/human_pika/orange_storage/episode_13/cam_left_wrist/frame_000015.jpg", "images/human_pika/orange_storage/episode_13/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`orange storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `orange storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -94.1, "hop_absolute_value": 94.1, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_pants_episode_21-train-bf_000330-af_000195-minus-0000211", "task": "fold pants", "image": [ "images/human_pika/fold_pants/episode_3/cam_high/frame_000000.jpg", "images/human_pika/fold_pants/episode_3/cam_high/frame_000548.jpg", "images/human_pika/fold_pants/episode_21/cam_high/frame_000330.jpg", "images/human_pika/fold_pants/episode_21/cam_left_wrist/frame_000330.jpg", "images/human_pika/fold_pants/episode_21/cam_right_wrist/frame_000330.jpg", "images/human_pika/fold_pants/episode_21/cam_high/frame_000195.jpg", "images/human_pika/fold_pants/episode_21/cam_left_wrist/frame_000195.jpg", "images/human_pika/fold_pants/episode_21/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold pants`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold pants`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.9, "hop_absolute_value": 40.9, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_test_tube_to_the_shelf_episode_16-train-bf_000510-af_000270-minus-0000201", "task": "put test tube to the shelf", "image": [ "images/human_pika/put_test_tube_to_the_shelf/episode_12/cam_high/frame_000000.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_12/cam_high/frame_000470.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_16/cam_high/frame_000510.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_16/cam_left_wrist/frame_000510.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_16/cam_right_wrist/frame_000510.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_16/cam_high/frame_000270.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_16/cam_left_wrist/frame_000270.jpg", "images/human_pika/put_test_tube_to_the_shelf/episode_16/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put test tube to the shelf`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put test tube to the shelf`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -47.1, "hop_absolute_value": 47.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_wipe_stains_episode_5-train-bf_000255-af_000210-minus-0000265", "task": "wipe stains", "image": [ "images/human_pika/wipe_stains/episode_24/cam_high/frame_000000.jpg", "images/human_pika/wipe_stains/episode_24/cam_high/frame_000608.jpg", "images/human_pika/wipe_stains/episode_5/cam_high/frame_000255.jpg", "images/human_pika/wipe_stains/episode_5/cam_left_wrist/frame_000255.jpg", "images/human_pika/wipe_stains/episode_5/cam_right_wrist/frame_000255.jpg", "images/human_pika/wipe_stains/episode_5/cam_high/frame_000210.jpg", "images/human_pika/wipe_stains/episode_5/cam_left_wrist/frame_000210.jpg", "images/human_pika/wipe_stains/episode_5/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`wipe stains`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `wipe stains`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -17.6, "hop_absolute_value": 17.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_cup_water_containing_episode_1-train-bf_000480-af_000270-minus-0000943", "task": "pour cup water containing", "image": [ "images/human_pika/pour_cup_water_containing/episode_24/cam_high/frame_000000.jpg", "images/human_pika/pour_cup_water_containing/episode_24/cam_high/frame_000514.jpg", "images/human_pika/pour_cup_water_containing/episode_1/cam_high/frame_000480.jpg", "images/human_pika/pour_cup_water_containing/episode_1/cam_left_wrist/frame_000480.jpg", "images/human_pika/pour_cup_water_containing/episode_1/cam_right_wrist/frame_000480.jpg", "images/human_pika/pour_cup_water_containing/episode_1/cam_high/frame_000270.jpg", "images/human_pika/pour_cup_water_containing/episode_1/cam_left_wrist/frame_000270.jpg", "images/human_pika/pour_cup_water_containing/episode_1/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour cup water containing`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour cup water containing`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -43.8, "hop_absolute_value": 43.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_towel_episode_15-train-bf_000405-af_000105-minus-0000856", "task": "fold towel", "image": [ "images/human_pika/fold_towel/episode_21/cam_high/frame_000000.jpg", "images/human_pika/fold_towel/episode_21/cam_high/frame_000495.jpg", "images/human_pika/fold_towel/episode_15/cam_high/frame_000405.jpg", "images/human_pika/fold_towel/episode_15/cam_left_wrist/frame_000405.jpg", "images/human_pika/fold_towel/episode_15/cam_right_wrist/frame_000405.jpg", "images/human_pika/fold_towel/episode_15/cam_high/frame_000105.jpg", "images/human_pika/fold_towel/episode_15/cam_left_wrist/frame_000105.jpg", "images/human_pika/fold_towel/episode_15/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold towel`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold towel`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -74.1, "hop_absolute_value": 74.1, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_11-train-bf_000270-af_000120-minus-0000168", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_11/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_11/cam_high/frame_000554.jpg", "images/human_pika/pour_tea/episode_11/cam_high/frame_000270.jpg", "images/human_pika/pour_tea/episode_11/cam_left_wrist/frame_000270.jpg", "images/human_pika/pour_tea/episode_11/cam_right_wrist/frame_000270.jpg", "images/human_pika/pour_tea/episode_11/cam_high/frame_000120.jpg", "images/human_pika/pour_tea/episode_11/cam_left_wrist/frame_000120.jpg", "images/human_pika/pour_tea/episode_11/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -55.6, "hop_absolute_value": 55.6, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_orange_storage_episode_4-train-bf_000180-af_000150-minus-0000266", "task": "orange storage", "image": [ "images/human_pika/orange_storage/episode_4/cam_high/frame_000000.jpg", "images/human_pika/orange_storage/episode_4/cam_high/frame_000545.jpg", "images/human_pika/orange_storage/episode_4/cam_high/frame_000180.jpg", "images/human_pika/orange_storage/episode_4/cam_left_wrist/frame_000180.jpg", "images/human_pika/orange_storage/episode_4/cam_right_wrist/frame_000180.jpg", "images/human_pika/orange_storage/episode_4/cam_high/frame_000150.jpg", "images/human_pika/orange_storage/episode_4/cam_left_wrist/frame_000150.jpg", "images/human_pika/orange_storage/episode_4/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`orange storage`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `orange storage`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_24-train-bf_000345-af_000150-minus-0000178", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_6/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_6/cam_high/frame_000630.jpg", "images/human_pika/pour_tea/episode_24/cam_high/frame_000345.jpg", "images/human_pika/pour_tea/episode_24/cam_left_wrist/frame_000345.jpg", "images/human_pika/pour_tea/episode_24/cam_right_wrist/frame_000345.jpg", "images/human_pika/pour_tea/episode_24/cam_high/frame_000150.jpg", "images/human_pika/pour_tea/episode_24/cam_left_wrist/frame_000150.jpg", "images/human_pika/pour_tea/episode_24/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -56.5, "hop_absolute_value": 56.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_3-train-bf_000255-af_000240-minus-0000286", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_21/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_21/cam_high/frame_000664.jpg", "images/human_pika/pour_tea/episode_3/cam_high/frame_000255.jpg", "images/human_pika/pour_tea/episode_3/cam_left_wrist/frame_000255.jpg", "images/human_pika/pour_tea/episode_3/cam_right_wrist/frame_000255.jpg", "images/human_pika/pour_tea/episode_3/cam_high/frame_000240.jpg", "images/human_pika/pour_tea/episode_3/cam_left_wrist/frame_000240.jpg", "images/human_pika/pour_tea/episode_3/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.9, "hop_absolute_value": 5.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_basket_with_peach_episode_4-train-bf_000090-af_000075-minus-0000262", "task": "pour basket with peach", "image": [ "images/human_pika/pour_basket_with_peach/episode_22/cam_high/frame_000000.jpg", "images/human_pika/pour_basket_with_peach/episode_22/cam_high/frame_000498.jpg", "images/human_pika/pour_basket_with_peach/episode_4/cam_high/frame_000090.jpg", "images/human_pika/pour_basket_with_peach/episode_4/cam_left_wrist/frame_000090.jpg", "images/human_pika/pour_basket_with_peach/episode_4/cam_right_wrist/frame_000090.jpg", "images/human_pika/pour_basket_with_peach/episode_4/cam_high/frame_000075.jpg", "images/human_pika/pour_basket_with_peach/episode_4/cam_left_wrist/frame_000075.jpg", "images/human_pika/pour_basket_with_peach/episode_4/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour basket with peach`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour basket with peach`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_cup_water_containing_episode_8-train-bf_000435-af_000225-minus-0000196", "task": "pour cup water containing", "image": [ "images/human_pika/pour_cup_water_containing/episode_2/cam_high/frame_000000.jpg", "images/human_pika/pour_cup_water_containing/episode_2/cam_high/frame_000494.jpg", "images/human_pika/pour_cup_water_containing/episode_8/cam_high/frame_000435.jpg", "images/human_pika/pour_cup_water_containing/episode_8/cam_left_wrist/frame_000435.jpg", "images/human_pika/pour_cup_water_containing/episode_8/cam_right_wrist/frame_000435.jpg", "images/human_pika/pour_cup_water_containing/episode_8/cam_high/frame_000225.jpg", "images/human_pika/pour_cup_water_containing/episode_8/cam_left_wrist/frame_000225.jpg", "images/human_pika/pour_cup_water_containing/episode_8/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour cup water containing`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour cup water containing`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -48.3, "hop_absolute_value": 48.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_tea_episode_3-train-bf_000570-af_000390-minus-0000251", "task": "pour tea", "image": [ "images/human_pika/pour_tea/episode_16/cam_high/frame_000000.jpg", "images/human_pika/pour_tea/episode_16/cam_high/frame_000651.jpg", "images/human_pika/pour_tea/episode_3/cam_high/frame_000570.jpg", "images/human_pika/pour_tea/episode_3/cam_left_wrist/frame_000570.jpg", "images/human_pika/pour_tea/episode_3/cam_right_wrist/frame_000570.jpg", "images/human_pika/pour_tea/episode_3/cam_high/frame_000390.jpg", "images/human_pika/pour_tea/episode_3/cam_left_wrist/frame_000390.jpg", "images/human_pika/pour_tea/episode_3/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour tea`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour tea`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -31.6, "hop_absolute_value": 31.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bowl_episode_12-train-bf_000345-af_000180-minus-0000208", "task": "pour bowl", "image": [ "images/human_pika/pour_bowl/episode_13/cam_high/frame_000000.jpg", "images/human_pika/pour_bowl/episode_13/cam_high/frame_000961.jpg", "images/human_pika/pour_bowl/episode_12/cam_high/frame_000345.jpg", "images/human_pika/pour_bowl/episode_12/cam_left_wrist/frame_000345.jpg", "images/human_pika/pour_bowl/episode_12/cam_right_wrist/frame_000345.jpg", "images/human_pika/pour_bowl/episode_12/cam_high/frame_000180.jpg", "images/human_pika/pour_bowl/episode_12/cam_left_wrist/frame_000180.jpg", "images/human_pika/pour_bowl/episode_12/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -47.8, "hop_absolute_value": 47.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "human_pika" }, { "id": "human_pika_wipe_stains_episode_18-train-bf_000390-af_000375-minus-0000275", "task": "wipe stains", "image": [ "images/human_pika/wipe_stains/episode_18/cam_high/frame_000000.jpg", "images/human_pika/wipe_stains/episode_18/cam_high/frame_000478.jpg", "images/human_pika/wipe_stains/episode_18/cam_high/frame_000390.jpg", "images/human_pika/wipe_stains/episode_18/cam_left_wrist/frame_000390.jpg", "images/human_pika/wipe_stains/episode_18/cam_right_wrist/frame_000390.jpg", "images/human_pika/wipe_stains/episode_18/cam_high/frame_000375.jpg", "images/human_pika/wipe_stains/episode_18/cam_left_wrist/frame_000375.jpg", "images/human_pika/wipe_stains/episode_18/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`wipe stains`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `wipe stains`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.8, "hop_absolute_value": 3.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_stir_coffee_episode_5-train-bf_000015-af_000000-minus-0000002", "task": "stir coffee", "image": [ "images/human_pika/stir_coffee/episode_5/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_5/cam_high/frame_000575.jpg", "images/human_pika/stir_coffee/episode_5/cam_high/frame_000015.jpg", "images/human_pika/stir_coffee/episode_5/cam_left_wrist/frame_000015.jpg", "images/human_pika/stir_coffee/episode_5/cam_right_wrist/frame_000015.jpg", "images/human_pika/stir_coffee/episode_5/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_5/cam_left_wrist/frame_000000.jpg", "images/human_pika/stir_coffee/episode_5/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stir coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stir coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bread_to_basket_episode_5-train-bf_000390-af_000375-minus-0000565", "task": "pour bread to basket", "image": [ "images/human_pika/pour_bread_to_basket/episode_5/cam_high/frame_000000.jpg", "images/human_pika/pour_bread_to_basket/episode_5/cam_high/frame_000529.jpg", "images/human_pika/pour_bread_to_basket/episode_5/cam_high/frame_000390.jpg", "images/human_pika/pour_bread_to_basket/episode_5/cam_left_wrist/frame_000390.jpg", "images/human_pika/pour_bread_to_basket/episode_5/cam_right_wrist/frame_000390.jpg", "images/human_pika/pour_bread_to_basket/episode_5/cam_high/frame_000375.jpg", "images/human_pika/pour_bread_to_basket/episode_5/cam_left_wrist/frame_000375.jpg", "images/human_pika/pour_bread_to_basket/episode_5/cam_right_wrist/frame_000375.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bread to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bread to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.8, "hop_absolute_value": 3.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_place_test_tube_episode_24-train-bf_000510-af_000495-minus-0000567", "task": "place test tube", "image": [ "images/human_pika/place_test_tube/episode_24/cam_high/frame_000000.jpg", "images/human_pika/place_test_tube/episode_24/cam_high/frame_000510.jpg", "images/human_pika/place_test_tube/episode_24/cam_high/frame_000510.jpg", "images/human_pika/place_test_tube/episode_24/cam_left_wrist/frame_000510.jpg", "images/human_pika/place_test_tube/episode_24/cam_right_wrist/frame_000510.jpg", "images/human_pika/place_test_tube/episode_24/cam_high/frame_000495.jpg", "images/human_pika/place_test_tube/episode_24/cam_left_wrist/frame_000495.jpg", "images/human_pika/place_test_tube/episode_24/cam_right_wrist/frame_000495.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`place test tube`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `place test tube`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.9, "hop_absolute_value": 2.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_stir_coffee_episode_2-train-bf_000465-af_000000-minus-0000032", "task": "stir coffee", "image": [ "images/human_pika/stir_coffee/episode_6/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_6/cam_high/frame_000520.jpg", "images/human_pika/stir_coffee/episode_2/cam_high/frame_000465.jpg", "images/human_pika/stir_coffee/episode_2/cam_left_wrist/frame_000465.jpg", "images/human_pika/stir_coffee/episode_2/cam_right_wrist/frame_000465.jpg", "images/human_pika/stir_coffee/episode_2/cam_high/frame_000000.jpg", "images/human_pika/stir_coffee/episode_2/cam_left_wrist/frame_000000.jpg", "images/human_pika/stir_coffee/episode_2/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`stir coffee`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `stir coffee`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_peach_to_the_basket_episode_11-train-bf_000210-af_000195-minus-0000591", "task": "put peach to the basket", "image": [ "images/human_pika/put_peach_to_the_basket/episode_9/cam_high/frame_000000.jpg", "images/human_pika/put_peach_to_the_basket/episode_9/cam_high/frame_000479.jpg", "images/human_pika/put_peach_to_the_basket/episode_11/cam_high/frame_000210.jpg", "images/human_pika/put_peach_to_the_basket/episode_11/cam_left_wrist/frame_000210.jpg", "images/human_pika/put_peach_to_the_basket/episode_11/cam_right_wrist/frame_000210.jpg", "images/human_pika/put_peach_to_the_basket/episode_11/cam_high/frame_000195.jpg", "images/human_pika/put_peach_to_the_basket/episode_11/cam_left_wrist/frame_000195.jpg", "images/human_pika/put_peach_to_the_basket/episode_11/cam_right_wrist/frame_000195.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put peach to the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put peach to the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.1, "hop_absolute_value": 7.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bread_to_basket_episode_20-train-bf_000015-af_000000-minus-0000002", "task": "pour bread to basket", "image": [ "images/human_pika/pour_bread_to_basket/episode_19/cam_high/frame_000000.jpg", "images/human_pika/pour_bread_to_basket/episode_19/cam_high/frame_000500.jpg", "images/human_pika/pour_bread_to_basket/episode_20/cam_high/frame_000015.jpg", "images/human_pika/pour_bread_to_basket/episode_20/cam_left_wrist/frame_000015.jpg", "images/human_pika/pour_bread_to_basket/episode_20/cam_right_wrist/frame_000015.jpg", "images/human_pika/pour_bread_to_basket/episode_20/cam_high/frame_000000.jpg", "images/human_pika/pour_bread_to_basket/episode_20/cam_left_wrist/frame_000000.jpg", "images/human_pika/pour_bread_to_basket/episode_20/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bread to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bread to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "human_pika" }, { "id": "human_pika_pour_bread_to_basket_episode_14-train-bf_000591-af_000570-minus-0000288", "task": "pour bread to basket", "image": [ "images/human_pika/pour_bread_to_basket/episode_14/cam_high/frame_000000.jpg", "images/human_pika/pour_bread_to_basket/episode_14/cam_high/frame_000591.jpg", "images/human_pika/pour_bread_to_basket/episode_14/cam_high/frame_000591.jpg", "images/human_pika/pour_bread_to_basket/episode_14/cam_left_wrist/frame_000591.jpg", "images/human_pika/pour_bread_to_basket/episode_14/cam_right_wrist/frame_000591.jpg", "images/human_pika/pour_bread_to_basket/episode_14/cam_high/frame_000570.jpg", "images/human_pika/pour_bread_to_basket/episode_14/cam_left_wrist/frame_000570.jpg", "images/human_pika/pour_bread_to_basket/episode_14/cam_right_wrist/frame_000570.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pour bread to basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pour bread to basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.6, "hop_absolute_value": 2.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_put_peach_to_the_basket_episode_8-train-bf_000360-af_000345-minus-0000594", "task": "put peach to the basket", "image": [ "images/human_pika/put_peach_to_the_basket/episode_2/cam_high/frame_000000.jpg", "images/human_pika/put_peach_to_the_basket/episode_2/cam_high/frame_000541.jpg", "images/human_pika/put_peach_to_the_basket/episode_8/cam_high/frame_000360.jpg", "images/human_pika/put_peach_to_the_basket/episode_8/cam_left_wrist/frame_000360.jpg", "images/human_pika/put_peach_to_the_basket/episode_8/cam_right_wrist/frame_000360.jpg", "images/human_pika/put_peach_to_the_basket/episode_8/cam_high/frame_000345.jpg", "images/human_pika/put_peach_to_the_basket/episode_8/cam_left_wrist/frame_000345.jpg", "images/human_pika/put_peach_to_the_basket/episode_8/cam_right_wrist/frame_000345.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put peach to the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put peach to the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.2, "hop_absolute_value": 4.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "human_pika_fold_towel_episode_1-train-bf_000555-af_000540-minus-0000590", "task": "fold towel", "image": [ "images/human_pika/fold_towel/episode_23/cam_high/frame_000000.jpg", "images/human_pika/fold_towel/episode_23/cam_high/frame_000505.jpg", "images/human_pika/fold_towel/episode_1/cam_high/frame_000555.jpg", "images/human_pika/fold_towel/episode_1/cam_left_wrist/frame_000555.jpg", "images/human_pika/fold_towel/episode_1/cam_right_wrist/frame_000555.jpg", "images/human_pika/fold_towel/episode_1/cam_high/frame_000540.jpg", "images/human_pika/fold_towel/episode_1/cam_left_wrist/frame_000540.jpg", "images/human_pika/fold_towel/episode_1/cam_right_wrist/frame_000540.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`fold towel`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `fold towel`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.7, "hop_absolute_value": 2.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "human_pika" }, { "id": "libero_data_pick_up_the_butter_and_place_it_in_the_basket_episode_45-train-bf_000015-af_000035-plus-0000706", "task": "pick up the butter and place it in the basket", "image": [ "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_34/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_34/cam_high/frame_000163.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_45/cam_high/frame_000015.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_45/cam_left_wrist/frame_000015.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_45/cam_right_wrist/frame_000015.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_45/cam_high/frame_000035.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_45/cam_left_wrist/frame_000035.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_45/cam_right_wrist/frame_000035.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the butter and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the butter and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.5, "hop_absolute_value": 10.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_bowl_on_the_plate_episode_30-train-bf_000035-af_000050-plus-0000213", "task": "put the bowl on the plate", "image": [ "images/libero_data/put_the_bowl_on_the_plate/episode_20/cam_high/frame_000000.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_20/cam_high/frame_000094.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_30/cam_high/frame_000035.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_30/cam_left_wrist/frame_000035.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_30/cam_right_wrist/frame_000035.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_30/cam_high/frame_000050.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_30/cam_left_wrist/frame_000050.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_30/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the bowl on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the bowl on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 23.1, "hop_absolute_value": 23.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket_episode_14-train-bf_000120-af_000125-plus-0000597", "task": "put both the alphabet soup and the cream cheese box in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_1/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_1/cam_high/frame_000289.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_14/cam_high/frame_000120.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_14/cam_left_wrist/frame_000120.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_14/cam_right_wrist/frame_000120.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_14/cam_high/frame_000125.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_14/cam_left_wrist/frame_000125.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_14/cam_right_wrist/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the cream cheese box in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the cream cheese box in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.8, "hop_absolute_value": 3.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_salad_dressing_and_place_it_in_the_basket_episode_32-train-bf_000015-af_000060-plus-0000560", "task": "pick up the salad dressing and place it in the basket", "image": [ "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_36/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_36/cam_high/frame_000127.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_32/cam_high/frame_000015.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_32/cam_left_wrist/frame_000015.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_32/cam_right_wrist/frame_000015.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_32/cam_high/frame_000060.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_32/cam_left_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_32/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the salad dressing and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the salad dressing and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 42.9, "hop_absolute_value": 42.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy_episode_40-train-bf_000010-af_000050-plus-0000315", "task": "pick up the book and place it in the back compartment of the caddy", "image": [ "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_43/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_43/cam_high/frame_000243.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_40/cam_high/frame_000010.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_40/cam_left_wrist/frame_000010.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_40/cam_right_wrist/frame_000010.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_40/cam_high/frame_000050.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_40/cam_left_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_40/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the book and place it in the back compartment of the caddy`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the book and place it in the back compartment of the caddy`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 29.6, "hop_absolute_value": 29.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_bbq_sauce_and_place_it_in_the_basket_episode_9-train-bf_000000-af_000030-plus-0000656", "task": "pick up the bbq sauce and place it in the basket", "image": [ "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_23/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_23/cam_high/frame_000132.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_9/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_9/cam_left_wrist/frame_000000.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_9/cam_right_wrist/frame_000000.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_9/cam_high/frame_000030.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_9/cam_left_wrist/frame_000030.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_9/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the bbq sauce and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the bbq sauce and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.7, "hop_absolute_value": 20.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket_episode_37-train-bf_000030-af_000070-plus-0000874", "task": "put both the alphabet soup and the cream cheese box in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_33/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_33/cam_high/frame_000269.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_37/cam_high/frame_000030.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_37/cam_left_wrist/frame_000030.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_37/cam_right_wrist/frame_000030.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_37/cam_high/frame_000070.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_37/cam_left_wrist/frame_000070.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_37/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the cream cheese box in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the cream cheese box in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 19.5, "hop_absolute_value": 19.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate_episode_10-train-bf_000090-af_000110-plus-0000324", "task": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate/episode_30/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate/episode_30/cam_high/frame_000169.jpg", "images/libero_data/pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate/episode_10/cam_high/frame_000090.jpg", "images/libero_data/pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate/episode_10/cam_left_wrist/frame_000090.jpg", "images/libero_data/pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate/episode_10/cam_right_wrist/frame_000090.jpg", "images/libero_data/pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate/episode_10/cam_high/frame_000110.jpg", "images/libero_data/pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate/episode_10/cam_left_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate/episode_10/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 44.4, "hop_absolute_value": 44.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_wine_bottle_on_the_rack_episode_12-train-bf_000035-af_000075-plus-0000860", "task": "put the wine bottle on the rack", "image": [ "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_high/frame_000000.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_high/frame_000176.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_high/frame_000035.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_left_wrist/frame_000035.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_right_wrist/frame_000035.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_high/frame_000075.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_left_wrist/frame_000075.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the wine bottle on the rack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the wine bottle on the rack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 28.6, "hop_absolute_value": 28.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_middle_drawer_of_the_cabinet_episode_30-train-bf_000060-af_000125-plus-0000445", "task": "open the middle drawer of the cabinet", "image": [ "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_30/cam_high/frame_000000.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_30/cam_high/frame_000135.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_30/cam_high/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_30/cam_left_wrist/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_30/cam_right_wrist/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_30/cam_high/frame_000125.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_30/cam_left_wrist/frame_000125.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_30/cam_right_wrist/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the middle drawer of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the middle drawer of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 86.7, "hop_absolute_value": 86.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it_episode_11-train-bf_000190-af_000220-plus-0000524", "task": "put the black bowl in the bottom drawer of the cabinet and close it", "image": [ "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_18/cam_high/frame_000000.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_18/cam_high/frame_000243.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_11/cam_high/frame_000190.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_11/cam_left_wrist/frame_000190.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_11/cam_right_wrist/frame_000190.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_11/cam_high/frame_000220.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_11/cam_left_wrist/frame_000220.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_11/cam_right_wrist/frame_000220.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the black bowl in the bottom drawer of the cabinet and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the black bowl in the bottom drawer of the cabinet and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_5-train-bf_000060-af_000140-plus-0000483", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_5/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_5/cam_high/frame_000152.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_5/cam_high/frame_000060.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_5/cam_left_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_5/cam_right_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_5/cam_high/frame_000140.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_5/cam_left_wrist/frame_000140.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_5/cam_right_wrist/frame_000140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 88.9, "hop_absolute_value": 88.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_middle_drawer_of_the_cabinet_episode_2-train-bf_000060-af_000085-plus-0000692", "task": "open the middle drawer of the cabinet", "image": [ "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_16/cam_high/frame_000000.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_16/cam_high/frame_000184.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_2/cam_high/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_2/cam_left_wrist/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_2/cam_right_wrist/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_2/cam_high/frame_000085.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_2/cam_left_wrist/frame_000085.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_2/cam_right_wrist/frame_000085.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the middle drawer of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the middle drawer of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 27.8, "hop_absolute_value": 27.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_moka_pots_on_the_stove_episode_18-train-bf_000135-af_000290-plus-0000463", "task": "put both moka pots on the stove", "image": [ "images/libero_data/put_both_moka_pots_on_the_stove/episode_31/cam_high/frame_000000.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_31/cam_high/frame_000465.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_18/cam_high/frame_000135.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_18/cam_left_wrist/frame_000135.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_18/cam_right_wrist/frame_000135.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_18/cam_high/frame_000290.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_18/cam_left_wrist/frame_000290.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_18/cam_right_wrist/frame_000290.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both moka pots on the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both moka pots on the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 62.0, "hop_absolute_value": 62.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_push_the_plate_to_the_front_of_the_stove_episode_15-train-bf_000080-af_000120-plus-0000396", "task": "push the plate to the front of the stove", "image": [ "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_27/cam_high/frame_000000.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_27/cam_high/frame_000219.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_15/cam_high/frame_000080.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_15/cam_left_wrist/frame_000080.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_15/cam_right_wrist/frame_000080.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_15/cam_high/frame_000120.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_15/cam_left_wrist/frame_000120.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_15/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`push the plate to the front of the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `push the plate to the front of the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 72.7, "hop_absolute_value": 72.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_wine_bottle_on_the_rack_episode_45-train-bf_000025-af_000155-plus-0000493", "task": "put the wine bottle on the rack", "image": [ "images/libero_data/put_the_wine_bottle_on_the_rack/episode_43/cam_high/frame_000000.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_43/cam_high/frame_000178.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_45/cam_high/frame_000025.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_45/cam_left_wrist/frame_000025.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_45/cam_right_wrist/frame_000025.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_45/cam_high/frame_000155.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_45/cam_left_wrist/frame_000155.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_45/cam_right_wrist/frame_000155.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the wine bottle on the rack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the wine bottle on the rack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 81.2, "hop_absolute_value": 81.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_push_the_plate_to_the_front_of_the_stove_episode_30-train-bf_000010-af_000130-plus-0000509", "task": "push the plate to the front of the stove", "image": [ "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_34/cam_high/frame_000000.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_34/cam_high/frame_000153.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_30/cam_high/frame_000010.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_30/cam_left_wrist/frame_000010.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_30/cam_right_wrist/frame_000010.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_30/cam_high/frame_000130.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_30/cam_left_wrist/frame_000130.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_30/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`push the plate to the front of the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `push the plate to the front of the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 92.3, "hop_absolute_value": 92.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_tomato_sauce_and_place_it_in_the_basket_episode_22-train-bf_000120-af_000155-plus-0000412", "task": "pick up the tomato sauce and place it in the basket", "image": [ "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_30/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_30/cam_high/frame_000145.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_22/cam_high/frame_000120.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_22/cam_left_wrist/frame_000120.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_22/cam_right_wrist/frame_000120.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_22/cam_high/frame_000155.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_22/cam_left_wrist/frame_000155.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_22/cam_right_wrist/frame_000155.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the tomato sauce and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the tomato sauce and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 63.6, "hop_absolute_value": 63.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate_episode_25-train-bf_000000-af_000075-plus-0000341", "task": "put the white mug on the plate and put the chocolate pudding to the right of the plate", "image": [ "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_23/cam_high/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_23/cam_high/frame_000219.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_25/cam_high/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_25/cam_left_wrist/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_25/cam_right_wrist/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_25/cam_high/frame_000075.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_25/cam_left_wrist/frame_000075.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_25/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the white mug on the plate and put the chocolate pudding to the right of the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the white mug on the plate and put the chocolate pudding to the right of the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 32.6, "hop_absolute_value": 32.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate_episode_28-train-bf_000095-af_000110-plus-0000292", "task": "pick up the black bowl on the wooden cabinet and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_high/frame_000140.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_high/frame_000095.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_left_wrist/frame_000095.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_right_wrist/frame_000095.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_high/frame_000110.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_left_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the wooden cabinet and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the wooden cabinet and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_salad_dressing_and_place_it_in_the_basket_episode_33-train-bf_000005-af_000025-plus-0000502", "task": "pick up the salad dressing and place it in the basket", "image": [ "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_33/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_33/cam_high/frame_000123.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_33/cam_high/frame_000005.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_33/cam_left_wrist/frame_000005.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_33/cam_right_wrist/frame_000005.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_33/cam_high/frame_000025.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_33/cam_left_wrist/frame_000025.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_33/cam_right_wrist/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the salad dressing and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the salad dressing and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 17.4, "hop_absolute_value": 17.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_16-train-bf_000020-af_000155-plus-0000516", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_2/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_2/cam_high/frame_000150.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_16/cam_high/frame_000020.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_16/cam_left_wrist/frame_000020.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_16/cam_right_wrist/frame_000020.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_16/cam_high/frame_000155.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_16/cam_left_wrist/frame_000155.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_16/cam_right_wrist/frame_000155.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 90.0, "hop_absolute_value": 90.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_yellow_and_white_mug_in_the_microwave_and_close_it_episode_10-train-bf_000085-af_000240-plus-0000510", "task": "put the yellow and white mug in the microwave and close it", "image": [ "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_10/cam_high/frame_000000.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_10/cam_high/frame_000284.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_10/cam_high/frame_000085.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_10/cam_left_wrist/frame_000085.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_10/cam_right_wrist/frame_000085.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_10/cam_high/frame_000240.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_10/cam_left_wrist/frame_000240.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_10/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the yellow and white mug in the microwave and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the yellow and white mug in the microwave and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 79.5, "hop_absolute_value": 79.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_butter_and_place_it_in_the_basket_episode_22-train-bf_000045-af_000120-plus-0000870", "task": "pick up the butter and place it in the basket", "image": [ "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_26/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_26/cam_high/frame_000155.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_22/cam_high/frame_000045.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_22/cam_left_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_22/cam_right_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_22/cam_high/frame_000120.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_22/cam_left_wrist/frame_000120.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_22/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the butter and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the butter and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate_episode_36-train-bf_000025-af_000080-plus-0000357", "task": "pick up the black bowl on the stove and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_36/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_36/cam_high/frame_000147.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_36/cam_high/frame_000025.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_36/cam_left_wrist/frame_000025.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_36/cam_right_wrist/frame_000025.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_36/cam_high/frame_000080.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_36/cam_left_wrist/frame_000080.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_36/cam_right_wrist/frame_000080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the stove and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the stove and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 45.8, "hop_absolute_value": 45.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_cream_cheese_box_and_the_butter_in_the_basket_episode_47-train-bf_000120-af_000205-plus-0000438", "task": "put both the cream cheese box and the butter in the basket", "image": [ "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_3/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_3/cam_high/frame_000287.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_47/cam_high/frame_000120.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_47/cam_left_wrist/frame_000120.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_47/cam_right_wrist/frame_000120.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_47/cam_high/frame_000205.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_47/cam_left_wrist/frame_000205.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_47/cam_right_wrist/frame_000205.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the cream cheese box and the butter in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the cream cheese box and the butter in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 63.0, "hop_absolute_value": 63.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket_episode_20-train-bf_000205-af_000310-plus-0000499", "task": "put both the alphabet soup and the tomato sauce in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_20/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_20/cam_high/frame_000338.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_20/cam_high/frame_000205.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_20/cam_left_wrist/frame_000205.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_20/cam_right_wrist/frame_000205.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_20/cam_high/frame_000310.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_20/cam_left_wrist/frame_000310.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_20/cam_right_wrist/frame_000310.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the tomato sauce in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the tomato sauce in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 80.8, "hop_absolute_value": 80.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket_episode_26-train-bf_000165-af_000170-plus-0000612", "task": "put both the alphabet soup and the tomato sauce in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_49/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_49/cam_high/frame_000277.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_26/cam_high/frame_000165.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_26/cam_left_wrist/frame_000165.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_26/cam_right_wrist/frame_000165.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_26/cam_high/frame_000170.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_26/cam_left_wrist/frame_000170.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_26/cam_right_wrist/frame_000170.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the tomato sauce in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the tomato sauce in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.8, "hop_absolute_value": 3.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_middle_drawer_of_the_cabinet_episode_3-train-bf_000060-af_000095-plus-0000317", "task": "open the middle drawer of the cabinet", "image": [ "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_high/frame_000000.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_high/frame_000130.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_high/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_left_wrist/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_right_wrist/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_high/frame_000095.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_left_wrist/frame_000095.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_right_wrist/frame_000095.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the middle drawer of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the middle drawer of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate_episode_9-train-bf_000045-af_000095-plus-0000339", "task": "pick up the black bowl from table center and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_33/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_33/cam_high/frame_000119.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_9/cam_high/frame_000045.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_9/cam_left_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_9/cam_right_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_9/cam_high/frame_000095.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_9/cam_left_wrist/frame_000095.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_9/cam_right_wrist/frame_000095.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl from table center and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl from table center and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_36-train-bf_000085-af_000130-plus-0000921", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_high/frame_000161.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_high/frame_000085.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_left_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_right_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_high/frame_000130.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_left_wrist/frame_000130.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate_episode_10-train-bf_000005-af_000095-plus-0000409", "task": "pick up the black bowl on the stove and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_10/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_10/cam_high/frame_000139.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_10/cam_high/frame_000005.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_10/cam_left_wrist/frame_000005.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_10/cam_right_wrist/frame_000005.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_10/cam_high/frame_000095.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_10/cam_left_wrist/frame_000095.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_10/cam_right_wrist/frame_000095.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the stove and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the stove and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 69.2, "hop_absolute_value": 69.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate_episode_6-train-bf_000085-af_000100-plus-0000311", "task": "pick up the black bowl on the cookie box and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_6/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_6/cam_high/frame_000110.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_6/cam_high/frame_000085.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_6/cam_left_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_6/cam_right_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_6/cam_high/frame_000100.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_6/cam_left_wrist/frame_000100.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_6/cam_right_wrist/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the cookie box and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the cookie box and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_bowl_on_the_plate_episode_30-train-bf_000015-af_000095-plus-0000370", "task": "put the bowl on the plate", "image": [ "images/libero_data/put_the_bowl_on_the_plate/episode_24/cam_high/frame_000000.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_24/cam_high/frame_000088.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_30/cam_high/frame_000015.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_30/cam_left_wrist/frame_000015.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_30/cam_right_wrist/frame_000015.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_30/cam_high/frame_000095.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_30/cam_left_wrist/frame_000095.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_30/cam_right_wrist/frame_000095.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the bowl on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the bowl on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 94.1, "hop_absolute_value": 94.1, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_middle_drawer_of_the_cabinet_episode_8-train-bf_000070-af_000095-plus-0000300", "task": "open the middle drawer of the cabinet", "image": [ "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_29/cam_high/frame_000000.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_29/cam_high/frame_000145.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_8/cam_high/frame_000070.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_8/cam_left_wrist/frame_000070.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_8/cam_right_wrist/frame_000070.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_8/cam_high/frame_000095.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_8/cam_left_wrist/frame_000095.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_8/cam_right_wrist/frame_000095.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the middle drawer of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the middle drawer of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_alphabet_soup_and_place_it_in_the_basket_episode_18-train-bf_000060-af_000110-plus-0000823", "task": "pick up the alphabet soup and place it in the basket", "image": [ "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_18/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_18/cam_high/frame_000151.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_18/cam_high/frame_000060.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_18/cam_left_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_18/cam_right_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_18/cam_high/frame_000110.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_18/cam_left_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_18/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the alphabet soup and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the alphabet soup and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 55.6, "hop_absolute_value": 55.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_bbq_sauce_and_place_it_in_the_basket_episode_43-train-bf_000095-af_000100-plus-0000624", "task": "pick up the bbq sauce and place it in the basket", "image": [ "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_11/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_11/cam_high/frame_000148.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_43/cam_high/frame_000095.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_43/cam_left_wrist/frame_000095.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_43/cam_right_wrist/frame_000095.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_43/cam_high/frame_000100.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_43/cam_left_wrist/frame_000100.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_43/cam_right_wrist/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the bbq sauce and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the bbq sauce and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 6.2, "hop_absolute_value": 6.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_middle_drawer_of_the_cabinet_episode_18-train-bf_000060-af_000125-plus-0000467", "task": "open the middle drawer of the cabinet", "image": [ "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_16/cam_high/frame_000000.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_16/cam_high/frame_000184.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_18/cam_high/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_18/cam_left_wrist/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_18/cam_right_wrist/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_18/cam_high/frame_000125.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_18/cam_left_wrist/frame_000125.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_18/cam_right_wrist/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the middle drawer of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the middle drawer of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_push_the_plate_to_the_front_of_the_stove_episode_45-train-bf_000005-af_000110-plus-0000454", "task": "push the plate to the front of the stove", "image": [ "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_17/cam_high/frame_000000.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_17/cam_high/frame_000133.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_45/cam_high/frame_000005.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_45/cam_left_wrist/frame_000005.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_45/cam_right_wrist/frame_000005.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_45/cam_high/frame_000110.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_45/cam_left_wrist/frame_000110.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_45/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`push the plate to the front of the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `push the plate to the front of the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 84.0, "hop_absolute_value": 84.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate_episode_33-train-bf_000090-af_000260-plus-0000546", "task": "put the white mug on the plate and put the chocolate pudding to the right of the plate", "image": [ "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_48/cam_high/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_48/cam_high/frame_000280.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_33/cam_high/frame_000090.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_33/cam_left_wrist/frame_000090.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_33/cam_right_wrist/frame_000090.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_33/cam_high/frame_000260.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_33/cam_left_wrist/frame_000260.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_33/cam_right_wrist/frame_000260.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the white mug on the plate and put the chocolate pudding to the right of the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the white mug on the plate and put the chocolate pudding to the right of the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 91.9, "hop_absolute_value": 91.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_bowl_on_the_stove_episode_4-train-bf_000005-af_000060-plus-0000277", "task": "put the bowl on the stove", "image": [ "images/libero_data/put_the_bowl_on_the_stove/episode_4/cam_high/frame_000000.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_4/cam_high/frame_000103.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_4/cam_high/frame_000005.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_4/cam_left_wrist/frame_000005.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_4/cam_right_wrist/frame_000005.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_4/cam_high/frame_000060.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_4/cam_left_wrist/frame_000060.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_4/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the bowl on the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the bowl on the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.9, "hop_absolute_value": 57.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_cream_cheese_and_place_it_in_the_basket_episode_18-train-bf_000075-af_000090-plus-0000714", "task": "pick up the cream cheese and place it in the basket", "image": [ "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_37/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_37/cam_high/frame_000134.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_18/cam_high/frame_000075.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_18/cam_left_wrist/frame_000075.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_18/cam_right_wrist/frame_000075.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_18/cam_high/frame_000090.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_18/cam_left_wrist/frame_000090.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_18/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the cream cheese and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the cream cheese and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 15.0, "hop_absolute_value": 15.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_2-train-bf_000025-af_000100-plus-0000400", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_10/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_10/cam_high/frame_000151.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_2/cam_high/frame_000025.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_2/cam_left_wrist/frame_000025.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_2/cam_right_wrist/frame_000025.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_2/cam_high/frame_000100.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_2/cam_left_wrist/frame_000100.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_2/cam_right_wrist/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_19-train-bf_000080-af_000135-plus-0000428", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_19/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_19/cam_high/frame_000157.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_19/cam_high/frame_000080.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_19/cam_left_wrist/frame_000080.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_19/cam_right_wrist/frame_000080.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_19/cam_high/frame_000135.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_19/cam_left_wrist/frame_000135.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_19/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 73.3, "hop_absolute_value": 73.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_butter_and_place_it_in_the_basket_episode_32-train-bf_000005-af_000050-plus-0000315", "task": "pick up the butter and place it in the basket", "image": [ "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_29/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_29/cam_high/frame_000163.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_32/cam_high/frame_000005.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_32/cam_left_wrist/frame_000005.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_32/cam_right_wrist/frame_000005.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_32/cam_high/frame_000050.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_32/cam_left_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_32/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the butter and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the butter and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 31.0, "hop_absolute_value": 31.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_top_drawer_and_put_the_bowl_inside_episode_28-train-bf_000130-af_000210-plus-0000414", "task": "open the top drawer and put the bowl inside", "image": [ "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_28/cam_high/frame_000000.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_28/cam_high/frame_000269.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_28/cam_high/frame_000130.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_28/cam_left_wrist/frame_000130.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_28/cam_right_wrist/frame_000130.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_28/cam_high/frame_000210.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_28/cam_left_wrist/frame_000210.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_28/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the top drawer and put the bowl inside`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the top drawer and put the bowl inside`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 59.3, "hop_absolute_value": 59.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate_episode_28-train-bf_000135-af_000140-plus-0000476", "task": "pick up the black bowl on the wooden cabinet and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_high/frame_000141.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_high/frame_000135.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_left_wrist/frame_000135.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_right_wrist/frame_000135.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_high/frame_000140.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_left_wrist/frame_000140.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_right_wrist/frame_000140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the wooden cabinet and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the wooden cabinet and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy_episode_20-train-bf_000075-af_000180-plus-0000563", "task": "pick up the book and place it in the back compartment of the caddy", "image": [ "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_19/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_19/cam_high/frame_000164.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_20/cam_high/frame_000075.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_20/cam_left_wrist/frame_000075.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_20/cam_right_wrist/frame_000075.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_20/cam_high/frame_000180.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_20/cam_left_wrist/frame_000180.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_20/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the book and place it in the back compartment of the caddy`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the book and place it in the back compartment of the caddy`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 95.5, "hop_absolute_value": 95.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_milk_and_place_it_in_the_basket_episode_44-train-bf_000095-af_000120-plus-0000346", "task": "pick up the milk and place it in the basket", "image": [ "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_5/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_5/cam_high/frame_000174.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_44/cam_high/frame_000095.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_44/cam_left_wrist/frame_000095.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_44/cam_right_wrist/frame_000095.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_44/cam_high/frame_000120.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_44/cam_left_wrist/frame_000120.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_44/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the milk and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the milk and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate_episode_18-train-bf_000065-af_000095-plus-0000311", "task": "pick up the black bowl next to the cookie box and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_26/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_26/cam_high/frame_000121.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_18/cam_high/frame_000065.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_18/cam_left_wrist/frame_000065.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_18/cam_right_wrist/frame_000065.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_18/cam_high/frame_000095.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_18/cam_left_wrist/frame_000095.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_18/cam_right_wrist/frame_000095.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl next to the cookie box and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl next to the cookie box and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy_episode_26-train-bf_000075-af_000120-plus-0000931", "task": "pick up the book and place it in the back compartment of the caddy", "image": [ "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_44/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_44/cam_high/frame_000173.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_26/cam_high/frame_000075.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_26/cam_left_wrist/frame_000075.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_26/cam_right_wrist/frame_000075.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_26/cam_high/frame_000120.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_26/cam_left_wrist/frame_000120.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_26/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the book and place it in the back compartment of the caddy`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the book and place it in the back compartment of the caddy`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 47.4, "hop_absolute_value": 47.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_moka_pots_on_the_stove_episode_31-train-bf_000115-af_000400-plus-0000530", "task": "put both moka pots on the stove", "image": [ "images/libero_data/put_both_moka_pots_on_the_stove/episode_31/cam_high/frame_000000.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_31/cam_high/frame_000465.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_31/cam_high/frame_000115.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_31/cam_left_wrist/frame_000115.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_31/cam_right_wrist/frame_000115.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_31/cam_high/frame_000400.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_31/cam_left_wrist/frame_000400.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_31/cam_right_wrist/frame_000400.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both moka pots on the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both moka pots on the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 81.4, "hop_absolute_value": 81.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_turn_on_the_stove_episode_35-train-bf_000020-af_000045-plus-0000199", "task": "turn on the stove", "image": [ "images/libero_data/turn_on_the_stove/episode_35/cam_high/frame_000000.jpg", "images/libero_data/turn_on_the_stove/episode_35/cam_high/frame_000085.jpg", "images/libero_data/turn_on_the_stove/episode_35/cam_high/frame_000020.jpg", "images/libero_data/turn_on_the_stove/episode_35/cam_left_wrist/frame_000020.jpg", "images/libero_data/turn_on_the_stove/episode_35/cam_right_wrist/frame_000020.jpg", "images/libero_data/turn_on_the_stove/episode_35/cam_high/frame_000045.jpg", "images/libero_data/turn_on_the_stove/episode_35/cam_left_wrist/frame_000045.jpg", "images/libero_data/turn_on_the_stove/episode_35/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 38.5, "hop_absolute_value": 38.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_top_drawer_and_put_the_bowl_inside_episode_33-train-bf_000015-af_000020-plus-0000286", "task": "open the top drawer and put the bowl inside", "image": [ "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_41/cam_high/frame_000000.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_41/cam_high/frame_000177.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_33/cam_high/frame_000015.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_33/cam_left_wrist/frame_000015.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_33/cam_right_wrist/frame_000015.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_33/cam_high/frame_000020.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_33/cam_left_wrist/frame_000020.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_33/cam_right_wrist/frame_000020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the top drawer and put the bowl inside`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the top drawer and put the bowl inside`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 2.7, "hop_absolute_value": 2.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_ketchup_and_place_it_in_the_basket_episode_33-train-bf_000065-af_000090-plus-0000727", "task": "pick up the ketchup and place it in the basket", "image": [ "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_33/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_33/cam_high/frame_000161.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_33/cam_high/frame_000065.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_33/cam_left_wrist/frame_000065.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_33/cam_right_wrist/frame_000065.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_33/cam_high/frame_000090.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_33/cam_left_wrist/frame_000090.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_33/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the ketchup and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the ketchup and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 26.3, "hop_absolute_value": 26.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate_episode_28-train-bf_000025-af_000135-plus-0000504", "task": "pick up the black bowl on the wooden cabinet and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_32/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_32/cam_high/frame_000129.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_high/frame_000025.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_left_wrist/frame_000025.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_right_wrist/frame_000025.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_high/frame_000135.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_left_wrist/frame_000135.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_28/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the wooden cabinet and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the wooden cabinet and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 95.7, "hop_absolute_value": 95.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_butter_and_place_it_in_the_basket_episode_34-train-bf_000085-af_000090-plus-0000579", "task": "pick up the butter and place it in the basket", "image": [ "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_6/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_6/cam_high/frame_000147.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_34/cam_high/frame_000085.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_34/cam_left_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_34/cam_right_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_34/cam_high/frame_000090.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_34/cam_left_wrist/frame_000090.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_34/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the butter and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the butter and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 6.7, "hop_absolute_value": 6.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_alphabet_soup_and_place_it_in_the_basket_episode_46-train-bf_000035-af_000045-plus-0000660", "task": "pick up the alphabet soup and place it in the basket", "image": [ "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_17/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_17/cam_high/frame_000153.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_46/cam_high/frame_000035.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_46/cam_left_wrist/frame_000035.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_46/cam_right_wrist/frame_000035.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_46/cam_high/frame_000045.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_46/cam_left_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_46/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the alphabet soup and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the alphabet soup and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 6.7, "hop_absolute_value": 6.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_bbq_sauce_and_place_it_in_the_basket_episode_3-train-bf_000055-af_000075-plus-0000608", "task": "pick up the bbq sauce and place it in the basket", "image": [ "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_3/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_3/cam_high/frame_000135.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_3/cam_high/frame_000055.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_3/cam_left_wrist/frame_000055.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_3/cam_right_wrist/frame_000055.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_3/cam_high/frame_000075.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_3/cam_left_wrist/frame_000075.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_3/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the bbq sauce and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the bbq sauce and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket_episode_27-train-bf_000160-af_000170-plus-0000846", "task": "put both the alphabet soup and the cream cheese box in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_39/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_39/cam_high/frame_000280.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_27/cam_high/frame_000160.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_27/cam_left_wrist/frame_000160.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_27/cam_right_wrist/frame_000160.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_27/cam_high/frame_000170.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_27/cam_left_wrist/frame_000170.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_27/cam_right_wrist/frame_000170.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the cream cheese box in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the cream cheese box in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 8.3, "hop_absolute_value": 8.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate_episode_48-train-bf_000165-af_000240-plus-0000440", "task": "put the white mug on the plate and put the chocolate pudding to the right of the plate", "image": [ "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_43/cam_high/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_43/cam_high/frame_000341.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_48/cam_high/frame_000165.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_48/cam_left_wrist/frame_000165.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_48/cam_right_wrist/frame_000165.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_48/cam_high/frame_000240.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_48/cam_left_wrist/frame_000240.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_48/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the white mug on the plate and put the chocolate pudding to the right of the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the white mug on the plate and put the chocolate pudding to the right of the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 65.2, "hop_absolute_value": 65.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_bowl_on_the_stove_episode_10-train-bf_000055-af_000070-plus-0000217", "task": "put the bowl on the stove", "image": [ "images/libero_data/put_the_bowl_on_the_stove/episode_8/cam_high/frame_000000.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_8/cam_high/frame_000100.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_10/cam_high/frame_000055.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_10/cam_left_wrist/frame_000055.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_10/cam_right_wrist/frame_000055.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_10/cam_high/frame_000070.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_10/cam_left_wrist/frame_000070.jpg", "images/libero_data/put_the_bowl_on_the_stove/episode_10/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the bowl on the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the bowl on the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 37.5, "hop_absolute_value": 37.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket_episode_45-train-bf_000005-af_000040-plus-0000868", "task": "put both the alphabet soup and the cream cheese box in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_20/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_20/cam_high/frame_000288.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_45/cam_high/frame_000005.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_45/cam_left_wrist/frame_000005.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_45/cam_right_wrist/frame_000005.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_45/cam_high/frame_000040.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_45/cam_left_wrist/frame_000040.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_45/cam_right_wrist/frame_000040.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the cream cheese box in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the cream cheese box in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.9, "hop_absolute_value": 11.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate_episode_40-train-bf_000035-af_000060-plus-0000216", "task": "pick up the black bowl between the plate and the ramekin and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_40/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_40/cam_high/frame_000094.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_40/cam_high/frame_000035.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_40/cam_left_wrist/frame_000035.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_40/cam_right_wrist/frame_000035.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_40/cam_high/frame_000060.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_40/cam_left_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_40/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl between the plate and the ramekin and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl between the plate and the ramekin and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 45.5, "hop_absolute_value": 45.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_top_drawer_and_put_the_bowl_inside_episode_10-train-bf_000100-af_000120-plus-0000911", "task": "open the top drawer and put the bowl inside", "image": [ "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_10/cam_high/frame_000000.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_10/cam_high/frame_000298.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_10/cam_high/frame_000100.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_10/cam_left_wrist/frame_000100.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_10/cam_right_wrist/frame_000100.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_10/cam_high/frame_000120.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_10/cam_left_wrist/frame_000120.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_10/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the top drawer and put the bowl inside`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the top drawer and put the bowl inside`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.3, "hop_absolute_value": 10.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it_episode_34-train-bf_000235-af_000280-plus-0000530", "task": "put the black bowl in the bottom drawer of the cabinet and close it", "image": [ "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_24/cam_high/frame_000000.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_24/cam_high/frame_000220.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_34/cam_high/frame_000235.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_34/cam_left_wrist/frame_000235.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_34/cam_right_wrist/frame_000235.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_34/cam_high/frame_000280.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_34/cam_left_wrist/frame_000280.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_34/cam_right_wrist/frame_000280.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the black bowl in the bottom drawer of the cabinet and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the black bowl in the bottom drawer of the cabinet and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 90.0, "hop_absolute_value": 90.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate_episode_16-train-bf_000050-af_000065-plus-0000582", "task": "pick up the black bowl on the wooden cabinet and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_38/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_38/cam_high/frame_000139.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_high/frame_000050.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_left_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_right_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_high/frame_000065.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_left_wrist/frame_000065.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_right_wrist/frame_000065.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the wooden cabinet and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the wooden cabinet and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_butter_and_place_it_in_the_basket_episode_16-train-bf_000045-af_000065-plus-0000291", "task": "pick up the butter and place it in the basket", "image": [ "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_16/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_16/cam_high/frame_000146.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_16/cam_high/frame_000045.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_16/cam_left_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_16/cam_right_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_16/cam_high/frame_000065.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_16/cam_left_wrist/frame_000065.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_16/cam_right_wrist/frame_000065.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the butter and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the butter and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate_episode_12-train-bf_000040-af_000085-plus-0000319", "task": "pick up the black bowl next to the ramekin and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_31/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_31/cam_high/frame_000119.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_12/cam_high/frame_000040.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_12/cam_left_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_12/cam_right_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_12/cam_high/frame_000085.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_12/cam_left_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_12/cam_right_wrist/frame_000085.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl next to the ramekin and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl next to the ramekin and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 64.3, "hop_absolute_value": 64.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_bbq_sauce_and_place_it_in_the_basket_episode_13-train-bf_000045-af_000075-plus-0000856", "task": "pick up the bbq sauce and place it in the basket", "image": [ "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_13/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_13/cam_high/frame_000220.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_13/cam_high/frame_000045.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_13/cam_left_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_13/cam_right_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_13/cam_high/frame_000075.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_13/cam_left_wrist/frame_000075.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_13/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the bbq sauce and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the bbq sauce and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 17.1, "hop_absolute_value": 17.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate_episode_19-train-bf_000010-af_000050-plus-0000224", "task": "pick up the black bowl between the plate and the ramekin and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_5/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_5/cam_high/frame_000133.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_19/cam_high/frame_000010.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_19/cam_left_wrist/frame_000010.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_19/cam_right_wrist/frame_000010.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_19/cam_high/frame_000050.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_19/cam_left_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_19/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl between the plate and the ramekin and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl between the plate and the ramekin and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate_episode_39-train-bf_000020-af_000060-plus-0000235", "task": "pick up the black bowl on the cookie box and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_16/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_16/cam_high/frame_000100.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_39/cam_high/frame_000020.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_39/cam_left_wrist/frame_000020.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_39/cam_right_wrist/frame_000020.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_39/cam_high/frame_000060.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_39/cam_left_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_39/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the cookie box and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the cookie box and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.1, "hop_absolute_value": 57.1, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_cream_cheese_box_and_the_butter_in_the_basket_episode_16-train-bf_000005-af_000185-plus-0000426", "task": "put both the cream cheese box and the butter in the basket", "image": [ "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_16/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_16/cam_high/frame_000320.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_16/cam_high/frame_000005.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_16/cam_left_wrist/frame_000005.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_16/cam_right_wrist/frame_000005.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_16/cam_high/frame_000185.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_16/cam_left_wrist/frame_000185.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_16/cam_right_wrist/frame_000185.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the cream cheese box and the butter in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the cream cheese box and the butter in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.1, "hop_absolute_value": 57.1, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_butter_and_place_it_in_the_basket_episode_14-train-bf_000000-af_000020-plus-0000639", "task": "pick up the butter and place it in the basket", "image": [ "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_14/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_14/cam_high/frame_000153.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_14/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_14/cam_left_wrist/frame_000000.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_14/cam_right_wrist/frame_000000.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_14/cam_high/frame_000020.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_14/cam_left_wrist/frame_000020.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_14/cam_right_wrist/frame_000020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the butter and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the butter and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 13.3, "hop_absolute_value": 13.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_cream_cheese_and_place_it_in_the_basket_episode_31-train-bf_000030-af_000130-plus-0000479", "task": "pick up the cream cheese and place it in the basket", "image": [ "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_39/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_39/cam_high/frame_000163.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_31/cam_high/frame_000030.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_31/cam_left_wrist/frame_000030.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_31/cam_right_wrist/frame_000030.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_31/cam_high/frame_000130.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_31/cam_left_wrist/frame_000130.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_31/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the cream cheese and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the cream cheese and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 87.0, "hop_absolute_value": 87.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_41-train-bf_000040-af_000060-plus-0000611", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_24/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_24/cam_high/frame_000167.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_41/cam_high/frame_000040.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_41/cam_left_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_41/cam_right_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_41/cam_high/frame_000060.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_41/cam_left_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_41/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 19.0, "hop_absolute_value": 19.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_push_the_plate_to_the_front_of_the_stove_episode_27-train-bf_000115-af_000125-plus-0000740", "task": "push the plate to the front of the stove", "image": [ "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_15/cam_high/frame_000000.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_15/cam_high/frame_000135.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_27/cam_high/frame_000115.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_27/cam_left_wrist/frame_000115.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_27/cam_right_wrist/frame_000115.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_27/cam_high/frame_000125.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_27/cam_left_wrist/frame_000125.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_27/cam_right_wrist/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`push the plate to the front of the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `push the plate to the front of the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate_episode_32-train-bf_000085-af_000231-plus-0000573", "task": "put the white mug on the plate and put the chocolate pudding to the right of the plate", "image": [ "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_16/cam_high/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_16/cam_high/frame_000283.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_32/cam_high/frame_000085.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_32/cam_left_wrist/frame_000085.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_32/cam_right_wrist/frame_000085.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_32/cam_high/frame_000231.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_32/cam_left_wrist/frame_000231.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_32/cam_right_wrist/frame_000231.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the white mug on the plate and put the chocolate pudding to the right of the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the white mug on the plate and put the chocolate pudding to the right of the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_wine_bottle_on_the_rack_episode_43-train-bf_000065-af_000105-plus-0000906", "task": "put the wine bottle on the rack", "image": [ "images/libero_data/put_the_wine_bottle_on_the_rack/episode_30/cam_high/frame_000000.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_30/cam_high/frame_000154.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_43/cam_high/frame_000065.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_43/cam_left_wrist/frame_000065.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_43/cam_right_wrist/frame_000065.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_43/cam_high/frame_000105.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_43/cam_left_wrist/frame_000105.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_43/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the wine bottle on the rack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the wine bottle on the rack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 36.4, "hop_absolute_value": 36.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_wine_bottle_on_top_of_the_cabinet_episode_26-train-bf_000030-af_000085-plus-0000323", "task": "put the wine bottle on top of the cabinet", "image": [ "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_25/cam_high/frame_000000.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_25/cam_high/frame_000099.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_26/cam_high/frame_000030.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_26/cam_left_wrist/frame_000030.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_26/cam_right_wrist/frame_000030.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_26/cam_high/frame_000085.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_26/cam_left_wrist/frame_000085.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_26/cam_right_wrist/frame_000085.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the wine bottle on top of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the wine bottle on top of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 78.6, "hop_absolute_value": 78.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_yellow_and_white_mug_in_the_microwave_and_close_it_episode_16-train-bf_000180-af_000250-plus-0000507", "task": "put the yellow and white mug in the microwave and close it", "image": [ "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_37/cam_high/frame_000000.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_37/cam_high/frame_000277.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_16/cam_high/frame_000180.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_16/cam_left_wrist/frame_000180.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_16/cam_right_wrist/frame_000180.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_16/cam_high/frame_000250.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_16/cam_left_wrist/frame_000250.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_16/cam_right_wrist/frame_000250.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the yellow and white mug in the microwave and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the yellow and white mug in the microwave and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 82.4, "hop_absolute_value": 82.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate_episode_24-train-bf_000090-af_000115-plus-0000357", "task": "pick up the black bowl next to the cookie box and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_25/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_25/cam_high/frame_000145.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_24/cam_high/frame_000090.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_24/cam_left_wrist/frame_000090.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_24/cam_right_wrist/frame_000090.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_24/cam_high/frame_000115.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_24/cam_left_wrist/frame_000115.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_24/cam_right_wrist/frame_000115.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl next to the cookie box and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl next to the cookie box and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_ketchup_and_place_it_in_the_basket_episode_41-train-bf_000085-af_000140-plus-0000478", "task": "pick up the ketchup and place it in the basket", "image": [ "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_41/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_41/cam_high/frame_000154.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_41/cam_high/frame_000085.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_41/cam_left_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_41/cam_right_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_41/cam_high/frame_000140.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_41/cam_left_wrist/frame_000140.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_41/cam_right_wrist/frame_000140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the ketchup and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the ketchup and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 84.6, "hop_absolute_value": 84.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_wine_bottle_on_the_rack_episode_12-train-bf_000105-af_000120-plus-0000783", "task": "put the wine bottle on the rack", "image": [ "images/libero_data/put_the_wine_bottle_on_the_rack/episode_6/cam_high/frame_000000.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_6/cam_high/frame_000149.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_high/frame_000105.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_left_wrist/frame_000105.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_right_wrist/frame_000105.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_high/frame_000120.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_left_wrist/frame_000120.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_12/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the wine bottle on the rack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the wine bottle on the rack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 21.4, "hop_absolute_value": 21.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate_episode_15-train-bf_000040-af_000080-plus-0000329", "task": "pick up the black bowl from table center and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_15/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_15/cam_high/frame_000133.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_15/cam_high/frame_000040.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_15/cam_left_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_15/cam_right_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_15/cam_high/frame_000080.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_15/cam_left_wrist/frame_000080.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_15/cam_right_wrist/frame_000080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl from table center and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl from table center and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 44.4, "hop_absolute_value": 44.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_alphabet_soup_and_place_it_in_the_basket_episode_38-train-bf_000040-af_000105-plus-0000385", "task": "pick up the alphabet soup and place it in the basket", "image": [ "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_26/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_26/cam_high/frame_000135.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_38/cam_high/frame_000040.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_38/cam_left_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_38/cam_right_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_38/cam_high/frame_000105.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_38/cam_left_wrist/frame_000105.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_38/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the alphabet soup and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the alphabet soup and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 56.5, "hop_absolute_value": 56.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_wine_bottle_on_the_rack_episode_41-train-bf_000115-af_000145-plus-0000409", "task": "put the wine bottle on the rack", "image": [ "images/libero_data/put_the_wine_bottle_on_the_rack/episode_44/cam_high/frame_000000.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_44/cam_high/frame_000172.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_41/cam_high/frame_000115.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_41/cam_left_wrist/frame_000115.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_41/cam_right_wrist/frame_000115.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_41/cam_high/frame_000145.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_41/cam_left_wrist/frame_000145.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_41/cam_right_wrist/frame_000145.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the wine bottle on the rack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the wine bottle on the rack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate_episode_29-train-bf_000050-af_000075-plus-0000550", "task": "pick up the black bowl between the plate and the ramekin and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_41/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_41/cam_high/frame_000094.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_29/cam_high/frame_000050.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_29/cam_left_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_29/cam_right_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_29/cam_high/frame_000075.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_29/cam_left_wrist/frame_000075.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_29/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl between the plate and the ramekin and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl between the plate and the ramekin and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 35.7, "hop_absolute_value": 35.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_orange_juice_and_place_it_in_the_basket_episode_21-train-bf_000000-af_000045-plus-0000300", "task": "pick up the orange juice and place it in the basket", "image": [ "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_21/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_21/cam_high/frame_000135.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_21/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_21/cam_left_wrist/frame_000000.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_21/cam_right_wrist/frame_000000.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_21/cam_high/frame_000045.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_21/cam_left_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_21/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the orange juice and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the orange juice and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_butter_and_place_it_in_the_basket_episode_28-train-bf_000015-af_000145-plus-0000914", "task": "pick up the butter and place it in the basket", "image": [ "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_37/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_37/cam_high/frame_000157.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_28/cam_high/frame_000015.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_28/cam_left_wrist/frame_000015.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_28/cam_right_wrist/frame_000015.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_28/cam_high/frame_000145.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_28/cam_left_wrist/frame_000145.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_28/cam_right_wrist/frame_000145.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the butter and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the butter and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 96.3, "hop_absolute_value": 96.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_cream_cheese_and_place_it_in_the_basket_episode_19-train-bf_000030-af_000126-plus-0000471", "task": "pick up the cream cheese and place it in the basket", "image": [ "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_31/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_31/cam_high/frame_000145.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_19/cam_high/frame_000030.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_19/cam_left_wrist/frame_000030.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_19/cam_right_wrist/frame_000030.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_19/cam_high/frame_000126.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_19/cam_left_wrist/frame_000126.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_19/cam_right_wrist/frame_000126.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the cream cheese and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the cream cheese and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket_episode_20-train-bf_000135-af_000150-plus-0000857", "task": "put both the alphabet soup and the cream cheese box in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_13/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_13/cam_high/frame_000263.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_20/cam_high/frame_000135.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_20/cam_left_wrist/frame_000135.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_20/cam_right_wrist/frame_000135.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_20/cam_high/frame_000150.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_20/cam_left_wrist/frame_000150.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_20/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the cream cheese box in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the cream cheese box in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_wine_bottle_on_the_rack_episode_40-train-bf_000045-af_000080-plus-0000770", "task": "put the wine bottle on the rack", "image": [ "images/libero_data/put_the_wine_bottle_on_the_rack/episode_18/cam_high/frame_000000.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_18/cam_high/frame_000183.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_40/cam_high/frame_000045.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_40/cam_left_wrist/frame_000045.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_40/cam_right_wrist/frame_000045.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_40/cam_high/frame_000080.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_40/cam_left_wrist/frame_000080.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_40/cam_right_wrist/frame_000080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the wine bottle on the rack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the wine bottle on the rack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 26.9, "hop_absolute_value": 26.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_bowl_on_top_of_the_cabinet_episode_41-train-bf_000030-af_000095-plus-0000360", "task": "put the bowl on top of the cabinet", "image": [ "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_49/cam_high/frame_000000.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_49/cam_high/frame_000097.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_41/cam_high/frame_000030.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_41/cam_left_wrist/frame_000030.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_41/cam_right_wrist/frame_000030.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_41/cam_high/frame_000095.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_41/cam_left_wrist/frame_000095.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_41/cam_right_wrist/frame_000095.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the bowl on top of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the bowl on top of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 92.9, "hop_absolute_value": 92.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_cream_cheese_box_and_the_butter_in_the_basket_episode_12-train-bf_000010-af_000130-plus-0000393", "task": "put both the cream cheese box and the butter in the basket", "image": [ "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_44/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_44/cam_high/frame_000275.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_12/cam_high/frame_000010.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_12/cam_left_wrist/frame_000010.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_12/cam_right_wrist/frame_000010.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_12/cam_high/frame_000130.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_12/cam_left_wrist/frame_000130.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_12/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the cream cheese box and the butter in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the cream cheese box and the butter in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate_episode_43-train-bf_000005-af_000025-plus-0000535", "task": "pick up the black bowl on the wooden cabinet and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_14/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_14/cam_high/frame_000135.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_43/cam_high/frame_000005.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_43/cam_left_wrist/frame_000005.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_43/cam_right_wrist/frame_000005.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_43/cam_high/frame_000025.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_43/cam_left_wrist/frame_000025.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_43/cam_right_wrist/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the wooden cabinet and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the wooden cabinet and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_next_to_the_plate_and_place_it_on_the_plate_episode_4-train-bf_000020-af_000090-plus-0000390", "task": "pick up the black bowl next to the plate and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_next_to_the_plate_and_place_it_on_the_plate/episode_20/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_plate_and_place_it_on_the_plate/episode_20/cam_high/frame_000124.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_plate_and_place_it_on_the_plate/episode_4/cam_high/frame_000020.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_plate_and_place_it_on_the_plate/episode_4/cam_left_wrist/frame_000020.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_plate_and_place_it_on_the_plate/episode_4/cam_right_wrist/frame_000020.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_plate_and_place_it_on_the_plate/episode_4/cam_high/frame_000090.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_plate_and_place_it_on_the_plate/episode_4/cam_left_wrist/frame_000090.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_plate_and_place_it_on_the_plate/episode_4/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl next to the plate and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl next to the plate and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.9, "hop_absolute_value": 60.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_turn_on_the_stove_episode_48-train-bf_000005-af_000035-plus-0000180", "task": "turn on the stove", "image": [ "images/libero_data/turn_on_the_stove/episode_39/cam_high/frame_000000.jpg", "images/libero_data/turn_on_the_stove/episode_39/cam_high/frame_000092.jpg", "images/libero_data/turn_on_the_stove/episode_48/cam_high/frame_000005.jpg", "images/libero_data/turn_on_the_stove/episode_48/cam_left_wrist/frame_000005.jpg", "images/libero_data/turn_on_the_stove/episode_48/cam_right_wrist/frame_000005.jpg", "images/libero_data/turn_on_the_stove/episode_48/cam_high/frame_000035.jpg", "images/libero_data/turn_on_the_stove/episode_48/cam_left_wrist/frame_000035.jpg", "images/libero_data/turn_on_the_stove/episode_48/cam_right_wrist/frame_000035.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate_episode_42-train-bf_000040-af_000045-plus-0000532", "task": "pick up the black bowl from table center and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_7/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_7/cam_high/frame_000122.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_42/cam_high/frame_000040.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_42/cam_left_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_42/cam_right_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_42/cam_high/frame_000045.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_42/cam_left_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_42/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl from table center and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl from table center and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.3, "hop_absolute_value": 5.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_tomato_sauce_and_place_it_in_the_basket_episode_49-train-bf_000115-af_000120-plus-0000613", "task": "pick up the tomato sauce and place it in the basket", "image": [ "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_6/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_6/cam_high/frame_000138.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_49/cam_high/frame_000115.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_49/cam_left_wrist/frame_000115.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_49/cam_right_wrist/frame_000115.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_49/cam_high/frame_000120.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_49/cam_left_wrist/frame_000120.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_49/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the tomato sauce and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the tomato sauce and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_turn_on_the_stove_and_put_the_moka_pot_on_it_episode_37-train-bf_000230-af_000225-minus-0000595", "task": "turn on the stove and put the moka pot on it", "image": [ "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_38/cam_high/frame_000000.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_38/cam_high/frame_000242.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_37/cam_high/frame_000230.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_37/cam_left_wrist/frame_000230.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_37/cam_right_wrist/frame_000230.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_37/cam_high/frame_000225.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_37/cam_left_wrist/frame_000225.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_37/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on the stove and put the moka pot on it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on the stove and put the moka pot on it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.2, "hop_absolute_value": 2.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate_episode_22-train-bf_000060-af_000050-minus-0000357", "task": "pick up the black bowl between the plate and the ramekin and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_36/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_36/cam_high/frame_000101.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_22/cam_high/frame_000060.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_22/cam_left_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_22/cam_right_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_22/cam_high/frame_000050.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_22/cam_left_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_22/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl between the plate and the ramekin and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl between the plate and the ramekin and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_bowl_on_the_plate_episode_46-train-bf_000065-af_000060-minus-0000112", "task": "put the bowl on the plate", "image": [ "images/libero_data/put_the_bowl_on_the_plate/episode_46/cam_high/frame_000000.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_46/cam_high/frame_000078.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_46/cam_high/frame_000065.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_46/cam_left_wrist/frame_000065.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_46/cam_right_wrist/frame_000065.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_46/cam_high/frame_000060.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_46/cam_left_wrist/frame_000060.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_46/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the bowl on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the bowl on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.7, "hop_absolute_value": 7.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it_episode_19-train-bf_000175-af_000075-minus-0000184", "task": "put the black bowl in the bottom drawer of the cabinet and close it", "image": [ "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_19/cam_high/frame_000000.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_19/cam_high/frame_000262.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_19/cam_high/frame_000175.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_19/cam_left_wrist/frame_000175.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_19/cam_right_wrist/frame_000175.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_19/cam_high/frame_000075.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_19/cam_left_wrist/frame_000075.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_19/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the black bowl in the bottom drawer of the cabinet and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the black bowl in the bottom drawer of the cabinet and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -57.1, "hop_absolute_value": 57.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_milk_and_place_it_in_the_basket_episode_23-train-bf_000140-af_000005-minus-0000033", "task": "pick up the milk and place it in the basket", "image": [ "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_5/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_5/cam_high/frame_000174.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_23/cam_high/frame_000140.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_23/cam_left_wrist/frame_000140.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_23/cam_right_wrist/frame_000140.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_23/cam_high/frame_000005.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_23/cam_left_wrist/frame_000005.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_23/cam_right_wrist/frame_000005.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the milk and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the milk and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -96.4, "hop_absolute_value": 96.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_cream_cheese_box_and_the_butter_in_the_basket_episode_30-train-bf_000020-af_000005-minus-0000117", "task": "put both the cream cheese box and the butter in the basket", "image": [ "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_30/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_30/cam_high/frame_000248.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_30/cam_high/frame_000020.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_30/cam_left_wrist/frame_000020.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_30/cam_right_wrist/frame_000020.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_30/cam_high/frame_000005.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_30/cam_left_wrist/frame_000005.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_30/cam_right_wrist/frame_000005.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the cream cheese box and the butter in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the cream cheese box and the butter in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_wine_bottle_on_top_of_the_cabinet_episode_14-train-bf_000070-af_000055-minus-0000144", "task": "put the wine bottle on top of the cabinet", "image": [ "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_19/cam_high/frame_000000.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_19/cam_high/frame_000113.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_14/cam_high/frame_000070.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_14/cam_left_wrist/frame_000070.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_14/cam_right_wrist/frame_000070.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_14/cam_high/frame_000055.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_14/cam_left_wrist/frame_000055.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_14/cam_right_wrist/frame_000055.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the wine bottle on top of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the wine bottle on top of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -21.4, "hop_absolute_value": 21.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate_episode_33-train-bf_000070-af_000020-minus-0000083", "task": "pick up the black bowl next to the ramekin and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_33/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_33/cam_high/frame_000116.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_33/cam_high/frame_000070.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_33/cam_left_wrist/frame_000070.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_33/cam_right_wrist/frame_000070.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_33/cam_high/frame_000020.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_33/cam_left_wrist/frame_000020.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_33/cam_right_wrist/frame_000020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl next to the ramekin and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl next to the ramekin and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_cream_cheese_and_place_it_in_the_basket_episode_1-train-bf_000180-af_000160-minus-0000634", "task": "pick up the cream cheese and place it in the basket", "image": [ "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_1/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_1/cam_high/frame_000186.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_1/cam_high/frame_000180.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_1/cam_left_wrist/frame_000180.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_1/cam_right_wrist/frame_000180.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_1/cam_high/frame_000160.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_1/cam_left_wrist/frame_000160.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_1/cam_right_wrist/frame_000160.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the cream cheese and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the cream cheese and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_middle_drawer_of_the_cabinet_episode_21-train-bf_000060-af_000025-minus-0000615", "task": "open the middle drawer of the cabinet", "image": [ "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_47/cam_high/frame_000000.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_47/cam_high/frame_000126.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_21/cam_high/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_21/cam_left_wrist/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_21/cam_right_wrist/frame_000060.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_21/cam_high/frame_000025.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_21/cam_left_wrist/frame_000025.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_21/cam_right_wrist/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the middle drawer of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the middle drawer of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -58.3, "hop_absolute_value": 58.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket_episode_5-train-bf_000050-af_000015-minus-0000116", "task": "put both the alphabet soup and the cream cheese box in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_48/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_48/cam_high/frame_000260.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_5/cam_high/frame_000050.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_5/cam_left_wrist/frame_000050.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_5/cam_right_wrist/frame_000050.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_5/cam_high/frame_000015.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_5/cam_left_wrist/frame_000015.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_5/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the cream cheese box in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the cream cheese box in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_36-train-bf_000110-af_000105-minus-0000566", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_8/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_8/cam_high/frame_000154.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_high/frame_000110.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_left_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_right_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_high/frame_000105.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_left_wrist/frame_000105.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_36/cam_right_wrist/frame_000105.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.5, "hop_absolute_value": 4.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate_episode_24-train-bf_000085-af_000005-minus-0000018", "task": "pick up the black bowl on the ramekin and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate/episode_22/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate/episode_22/cam_high/frame_000106.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate/episode_24/cam_high/frame_000085.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate/episode_24/cam_left_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate/episode_24/cam_right_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate/episode_24/cam_high/frame_000005.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate/episode_24/cam_left_wrist/frame_000005.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate/episode_24/cam_right_wrist/frame_000005.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the ramekin and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the ramekin and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -94.1, "hop_absolute_value": 94.1, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket_episode_3-train-bf_000255-af_000155-minus-0000237", "task": "put both the alphabet soup and the cream cheese box in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_22/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_22/cam_high/frame_000266.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_3/cam_high/frame_000255.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_3/cam_left_wrist/frame_000255.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_3/cam_right_wrist/frame_000255.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_3/cam_high/frame_000155.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_3/cam_left_wrist/frame_000155.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_3/cam_right_wrist/frame_000155.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the cream cheese box in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the cream cheese box in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -39.2, "hop_absolute_value": 39.2, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_cream_cheese_and_place_it_in_the_basket_episode_20-train-bf_000045-af_000015-minus-0000102", "task": "pick up the cream cheese and place it in the basket", "image": [ "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_23/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_23/cam_high/frame_000139.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_20/cam_high/frame_000045.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_20/cam_left_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_20/cam_right_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_20/cam_high/frame_000015.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_20/cam_left_wrist/frame_000015.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_20/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the cream cheese and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the cream cheese and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_tomato_sauce_and_place_it_in_the_basket_episode_38-train-bf_000040-af_000030-minus-0000244", "task": "pick up the tomato sauce and place it in the basket", "image": [ "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_43/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_43/cam_high/frame_000174.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_38/cam_high/frame_000040.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_38/cam_left_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_38/cam_right_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_38/cam_high/frame_000030.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_38/cam_left_wrist/frame_000030.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_38/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the tomato sauce and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the tomato sauce and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_moka_pots_on_the_stove_episode_22-train-bf_000260-af_000245-minus-0000868", "task": "put both moka pots on the stove", "image": [ "images/libero_data/put_both_moka_pots_on_the_stove/episode_40/cam_high/frame_000000.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_40/cam_high/frame_000415.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_22/cam_high/frame_000260.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_22/cam_left_wrist/frame_000260.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_22/cam_right_wrist/frame_000260.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_22/cam_high/frame_000245.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_22/cam_left_wrist/frame_000245.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_22/cam_right_wrist/frame_000245.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both moka pots on the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both moka pots on the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.8, "hop_absolute_value": 5.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_middle_drawer_of_the_cabinet_episode_25-train-bf_000145-af_000095-minus-0000237", "task": "open the middle drawer of the cabinet", "image": [ "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_25/cam_high/frame_000000.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_25/cam_high/frame_000159.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_25/cam_high/frame_000145.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_25/cam_left_wrist/frame_000145.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_25/cam_right_wrist/frame_000145.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_25/cam_high/frame_000095.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_25/cam_left_wrist/frame_000095.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_25/cam_right_wrist/frame_000095.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the middle drawer of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the middle drawer of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -34.5, "hop_absolute_value": 34.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket_episode_19-train-bf_000140-af_000130-minus-0000752", "task": "put both the alphabet soup and the cream cheese box in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_19/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_19/cam_high/frame_000311.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_19/cam_high/frame_000140.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_19/cam_left_wrist/frame_000140.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_19/cam_right_wrist/frame_000140.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_19/cam_high/frame_000130.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_19/cam_left_wrist/frame_000130.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_19/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the cream cheese box in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the cream cheese box in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.1, "hop_absolute_value": 7.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket_episode_19-train-bf_000285-af_000050-minus-0000106", "task": "put both the alphabet soup and the tomato sauce in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_11/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_11/cam_high/frame_000249.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_19/cam_high/frame_000285.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_19/cam_left_wrist/frame_000285.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_19/cam_right_wrist/frame_000285.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_19/cam_high/frame_000050.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_19/cam_left_wrist/frame_000050.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_19/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the tomato sauce in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the tomato sauce in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -82.5, "hop_absolute_value": 82.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate_episode_26-train-bf_000115-af_000065-minus-0000198", "task": "pick up the black bowl on the wooden cabinet and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_26/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_26/cam_high/frame_000142.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_26/cam_high/frame_000115.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_26/cam_left_wrist/frame_000115.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_26/cam_right_wrist/frame_000115.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_26/cam_high/frame_000065.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_26/cam_left_wrist/frame_000065.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_26/cam_right_wrist/frame_000065.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the wooden cabinet and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the wooden cabinet and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -43.5, "hop_absolute_value": 43.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_push_the_plate_to_the_front_of_the_stove_episode_0-train-bf_000040-af_000030-minus-0000650", "task": "push the plate to the front of the stove", "image": [ "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_10/cam_high/frame_000000.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_10/cam_high/frame_000214.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_0/cam_high/frame_000040.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_0/cam_left_wrist/frame_000040.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_0/cam_right_wrist/frame_000040.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_0/cam_high/frame_000030.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_0/cam_left_wrist/frame_000030.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_0/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`push the plate to the front of the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `push the plate to the front of the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_push_the_plate_to_the_front_of_the_stove_episode_25-train-bf_000060-af_000050-minus-0000603", "task": "push the plate to the front of the stove", "image": [ "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_4/cam_high/frame_000000.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_4/cam_high/frame_000192.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_25/cam_high/frame_000060.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_25/cam_left_wrist/frame_000060.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_25/cam_right_wrist/frame_000060.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_25/cam_high/frame_000050.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_25/cam_left_wrist/frame_000050.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_25/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`push the plate to the front of the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `push the plate to the front of the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate_episode_1-train-bf_000130-af_000030-minus-0000077", "task": "pick up the black bowl next to the cookie box and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_22/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_22/cam_high/frame_000123.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_1/cam_high/frame_000130.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_1/cam_left_wrist/frame_000130.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_1/cam_right_wrist/frame_000130.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_1/cam_high/frame_000030.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_1/cam_left_wrist/frame_000030.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_1/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl next to the cookie box and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl next to the cookie box and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -76.9, "hop_absolute_value": 76.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_14-train-bf_000065-af_000050-minus-0000245", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_15/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_15/cam_high/frame_000158.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_14/cam_high/frame_000065.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_14/cam_left_wrist/frame_000065.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_14/cam_right_wrist/frame_000065.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_14/cam_high/frame_000050.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_14/cam_left_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_14/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -23.1, "hop_absolute_value": 23.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_middle_drawer_of_the_cabinet_episode_47-train-bf_000075-af_000025-minus-0000112", "task": "open the middle drawer of the cabinet", "image": [ "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_34/cam_high/frame_000000.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_34/cam_high/frame_000141.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_47/cam_high/frame_000075.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_47/cam_left_wrist/frame_000075.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_47/cam_right_wrist/frame_000075.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_47/cam_high/frame_000025.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_47/cam_left_wrist/frame_000025.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_47/cam_right_wrist/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the middle drawer of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the middle drawer of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_bowl_on_the_plate_episode_7-train-bf_000055-af_000045-minus-0000149", "task": "put the bowl on the plate", "image": [ "images/libero_data/put_the_bowl_on_the_plate/episode_5/cam_high/frame_000000.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_5/cam_high/frame_000088.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_7/cam_high/frame_000055.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_7/cam_left_wrist/frame_000055.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_7/cam_right_wrist/frame_000055.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_7/cam_high/frame_000045.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_7/cam_left_wrist/frame_000045.jpg", "images/libero_data/put_the_bowl_on_the_plate/episode_7/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the bowl on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the bowl on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -18.2, "hop_absolute_value": 18.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_yellow_and_white_mug_in_the_microwave_and_close_it_episode_42-train-bf_000140-af_000135-minus-0000643", "task": "put the yellow and white mug in the microwave and close it", "image": [ "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_28/cam_high/frame_000000.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_28/cam_high/frame_000238.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_42/cam_high/frame_000140.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_42/cam_left_wrist/frame_000140.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_42/cam_right_wrist/frame_000140.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_42/cam_high/frame_000135.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_42/cam_left_wrist/frame_000135.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_42/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the yellow and white mug in the microwave and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the yellow and white mug in the microwave and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.6, "hop_absolute_value": 3.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_middle_drawer_of_the_cabinet_episode_3-train-bf_000120-af_000000-minus-0000032", "task": "open the middle drawer of the cabinet", "image": [ "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_41/cam_high/frame_000000.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_41/cam_high/frame_000135.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_high/frame_000120.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_left_wrist/frame_000120.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_right_wrist/frame_000120.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_high/frame_000000.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_left_wrist/frame_000000.jpg", "images/libero_data/open_the_middle_drawer_of_the_cabinet/episode_3/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the middle drawer of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the middle drawer of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_orange_juice_and_place_it_in_the_basket_episode_26-train-bf_000070-af_000035-minus-0000165", "task": "pick up the orange juice and place it in the basket", "image": [ "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_12/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_12/cam_high/frame_000143.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_26/cam_high/frame_000070.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_26/cam_left_wrist/frame_000070.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_26/cam_right_wrist/frame_000070.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_26/cam_high/frame_000035.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_26/cam_left_wrist/frame_000035.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_26/cam_right_wrist/frame_000035.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the orange juice and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the orange juice and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate_episode_22-train-bf_000236-af_000025-minus-0000069", "task": "put the white mug on the plate and put the chocolate pudding to the right of the plate", "image": [ "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_32/cam_high/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_32/cam_high/frame_000231.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_22/cam_high/frame_000236.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_22/cam_left_wrist/frame_000236.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_22/cam_right_wrist/frame_000236.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_22/cam_high/frame_000025.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_22/cam_left_wrist/frame_000025.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_22/cam_right_wrist/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the white mug on the plate and put the chocolate pudding to the right of the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the white mug on the plate and put the chocolate pudding to the right of the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -89.4, "hop_absolute_value": 89.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_moka_pots_on_the_stove_episode_37-train-bf_000340-af_000135-minus-0000167", "task": "put both moka pots on the stove", "image": [ "images/libero_data/put_both_moka_pots_on_the_stove/episode_37/cam_high/frame_000000.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_37/cam_high/frame_000403.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_37/cam_high/frame_000340.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_37/cam_left_wrist/frame_000340.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_37/cam_right_wrist/frame_000340.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_37/cam_high/frame_000135.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_37/cam_left_wrist/frame_000135.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_37/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both moka pots on the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both moka pots on the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -60.3, "hop_absolute_value": 60.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_top_drawer_and_put_the_bowl_inside_episode_16-train-bf_000195-af_000180-minus-0000290", "task": "open the top drawer and put the bowl inside", "image": [ "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_33/cam_high/frame_000000.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_33/cam_high/frame_000201.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_16/cam_high/frame_000195.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_16/cam_left_wrist/frame_000195.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_16/cam_right_wrist/frame_000195.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_16/cam_high/frame_000180.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_16/cam_left_wrist/frame_000180.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_16/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the top drawer and put the bowl inside`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the top drawer and put the bowl inside`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.7, "hop_absolute_value": 7.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_wine_bottle_on_the_rack_episode_21-train-bf_000152-af_000125-minus-0000252", "task": "put the wine bottle on the rack", "image": [ "images/libero_data/put_the_wine_bottle_on_the_rack/episode_13/cam_high/frame_000000.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_13/cam_high/frame_000154.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_21/cam_high/frame_000152.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_21/cam_left_wrist/frame_000152.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_21/cam_right_wrist/frame_000152.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_21/cam_high/frame_000125.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_21/cam_left_wrist/frame_000125.jpg", "images/libero_data/put_the_wine_bottle_on_the_rack/episode_21/cam_right_wrist/frame_000125.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the wine bottle on the rack`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the wine bottle on the rack`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_open_the_top_drawer_and_put_the_bowl_inside_episode_41-train-bf_000100-af_000070-minus-0000231", "task": "open the top drawer and put the bowl inside", "image": [ "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_49/cam_high/frame_000000.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_49/cam_high/frame_000208.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_41/cam_high/frame_000100.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_41/cam_left_wrist/frame_000100.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_41/cam_right_wrist/frame_000100.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_41/cam_high/frame_000070.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_41/cam_left_wrist/frame_000070.jpg", "images/libero_data/open_the_top_drawer_and_put_the_bowl_inside/episode_41/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open the top drawer and put the bowl inside`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open the top drawer and put the bowl inside`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -30.0, "hop_absolute_value": 30.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate_episode_14-train-bf_000040-af_000035-minus-0000801", "task": "put the white mug on the plate and put the chocolate pudding to the right of the plate", "image": [ "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_25/cam_high/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_25/cam_high/frame_000231.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_14/cam_high/frame_000040.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_14/cam_left_wrist/frame_000040.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_14/cam_right_wrist/frame_000040.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_14/cam_high/frame_000035.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_14/cam_left_wrist/frame_000035.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_14/cam_right_wrist/frame_000035.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the white mug on the plate and put the chocolate pudding to the right of the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the white mug on the plate and put the chocolate pudding to the right of the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate_episode_31-train-bf_000085-af_000010-minus-0000042", "task": "pick up the black bowl next to the ramekin and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_48/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_48/cam_high/frame_000140.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_31/cam_high/frame_000085.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_31/cam_left_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_31/cam_right_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_31/cam_high/frame_000010.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_31/cam_left_wrist/frame_000010.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_31/cam_right_wrist/frame_000010.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl next to the ramekin and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl next to the ramekin and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -88.2, "hop_absolute_value": 88.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_bowl_on_top_of_the_cabinet_episode_22-train-bf_000070-af_000015-minus-0000058", "task": "put the bowl on top of the cabinet", "image": [ "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_29/cam_high/frame_000000.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_29/cam_high/frame_000089.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_22/cam_high/frame_000070.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_22/cam_left_wrist/frame_000070.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_22/cam_right_wrist/frame_000070.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_22/cam_high/frame_000015.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_22/cam_left_wrist/frame_000015.jpg", "images/libero_data/put_the_bowl_on_top_of_the_cabinet/episode_22/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the bowl on top of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the bowl on top of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -78.6, "hop_absolute_value": 78.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_salad_dressing_and_place_it_in_the_basket_episode_44-train-bf_000085-af_000050-minus-0000172", "task": "pick up the salad dressing and place it in the basket", "image": [ "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_6/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_6/cam_high/frame_000125.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_44/cam_high/frame_000085.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_44/cam_left_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_44/cam_right_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_44/cam_high/frame_000050.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_44/cam_left_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_44/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the salad dressing and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the salad dressing and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -41.2, "hop_absolute_value": 41.2, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_wine_bottle_on_top_of_the_cabinet_episode_12-train-bf_000065-af_000005-minus-0000010", "task": "put the wine bottle on top of the cabinet", "image": [ "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_13/cam_high/frame_000000.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_13/cam_high/frame_000092.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_12/cam_high/frame_000065.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_12/cam_left_wrist/frame_000065.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_12/cam_right_wrist/frame_000065.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_12/cam_high/frame_000005.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_12/cam_left_wrist/frame_000005.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_12/cam_right_wrist/frame_000005.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the wine bottle on top of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the wine bottle on top of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -92.3, "hop_absolute_value": 92.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_yellow_and_white_mug_in_the_microwave_and_close_it_episode_24-train-bf_000135-af_000130-minus-0000636", "task": "put the yellow and white mug in the microwave and close it", "image": [ "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_23/cam_high/frame_000000.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_23/cam_high/frame_000342.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_24/cam_high/frame_000135.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_24/cam_left_wrist/frame_000135.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_24/cam_right_wrist/frame_000135.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_24/cam_high/frame_000130.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_24/cam_left_wrist/frame_000130.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_24/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the yellow and white mug in the microwave and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the yellow and white mug in the microwave and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.7, "hop_absolute_value": 3.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_moka_pots_on_the_stove_episode_44-train-bf_000310-af_000290-minus-0000756", "task": "put both moka pots on the stove", "image": [ "images/libero_data/put_both_moka_pots_on_the_stove/episode_15/cam_high/frame_000000.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_15/cam_high/frame_000372.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_44/cam_high/frame_000310.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_44/cam_left_wrist/frame_000310.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_44/cam_right_wrist/frame_000310.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_44/cam_high/frame_000290.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_44/cam_left_wrist/frame_000290.jpg", "images/libero_data/put_both_moka_pots_on_the_stove/episode_44/cam_right_wrist/frame_000290.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both moka pots on the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both moka pots on the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.5, "hop_absolute_value": 6.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate_episode_16-train-bf_000090-af_000070-minus-0000614", "task": "pick up the black bowl on the wooden cabinet and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_13/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_13/cam_high/frame_000133.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_high/frame_000090.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_left_wrist/frame_000090.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_right_wrist/frame_000090.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_high/frame_000070.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_left_wrist/frame_000070.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the wooden cabinet and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the wooden cabinet and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -22.2, "hop_absolute_value": 22.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_push_the_plate_to_the_front_of_the_stove_episode_25-train-bf_000130-af_000005-minus-0000027", "task": "push the plate to the front of the stove", "image": [ "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_35/cam_high/frame_000000.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_35/cam_high/frame_000168.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_25/cam_high/frame_000130.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_25/cam_left_wrist/frame_000130.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_25/cam_right_wrist/frame_000130.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_25/cam_high/frame_000005.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_25/cam_left_wrist/frame_000005.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_25/cam_right_wrist/frame_000005.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`push the plate to the front of the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `push the plate to the front of the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -96.2, "hop_absolute_value": 96.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it_episode_4-train-bf_000140-af_000015-minus-0000051", "task": "put the black bowl in the bottom drawer of the cabinet and close it", "image": [ "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_27/cam_high/frame_000000.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_27/cam_high/frame_000234.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_4/cam_high/frame_000140.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_4/cam_left_wrist/frame_000140.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_4/cam_right_wrist/frame_000140.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_4/cam_high/frame_000015.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_4/cam_left_wrist/frame_000015.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_4/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the black bowl in the bottom drawer of the cabinet and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the black bowl in the bottom drawer of the cabinet and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -89.3, "hop_absolute_value": 89.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_cream_cheese_in_the_bowl_episode_13-train-bf_000075-af_000020-minus-0000063", "task": "put the cream cheese in the bowl", "image": [ "images/libero_data/put_the_cream_cheese_in_the_bowl/episode_28/cam_high/frame_000000.jpg", "images/libero_data/put_the_cream_cheese_in_the_bowl/episode_28/cam_high/frame_000094.jpg", "images/libero_data/put_the_cream_cheese_in_the_bowl/episode_13/cam_high/frame_000075.jpg", "images/libero_data/put_the_cream_cheese_in_the_bowl/episode_13/cam_left_wrist/frame_000075.jpg", "images/libero_data/put_the_cream_cheese_in_the_bowl/episode_13/cam_right_wrist/frame_000075.jpg", "images/libero_data/put_the_cream_cheese_in_the_bowl/episode_13/cam_high/frame_000020.jpg", "images/libero_data/put_the_cream_cheese_in_the_bowl/episode_13/cam_left_wrist/frame_000020.jpg", "images/libero_data/put_the_cream_cheese_in_the_bowl/episode_13/cam_right_wrist/frame_000020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the cream cheese in the bowl`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the cream cheese in the bowl`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.3, "hop_absolute_value": 73.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_alphabet_soup_and_place_it_in_the_basket_episode_26-train-bf_000110-af_000085-minus-0000596", "task": "pick up the alphabet soup and place it in the basket", "image": [ "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_35/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_35/cam_high/frame_000149.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_26/cam_high/frame_000110.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_26/cam_left_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_26/cam_right_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_26/cam_high/frame_000085.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_26/cam_left_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_26/cam_right_wrist/frame_000085.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the alphabet soup and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the alphabet soup and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -22.7, "hop_absolute_value": 22.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_cream_cheese_and_place_it_in_the_basket_episode_49-train-bf_000050-af_000005-minus-0000035", "task": "pick up the cream cheese and place it in the basket", "image": [ "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_36/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_36/cam_high/frame_000137.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_49/cam_high/frame_000050.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_49/cam_left_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_49/cam_right_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_49/cam_high/frame_000005.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_49/cam_left_wrist/frame_000005.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_49/cam_right_wrist/frame_000005.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the cream cheese and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the cream cheese and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -90.0, "hop_absolute_value": 90.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate_episode_22-train-bf_000110-af_000050-minus-0000149", "task": "pick up the black bowl next to the ramekin and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_9/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_9/cam_high/frame_000126.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_22/cam_high/frame_000110.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_22/cam_left_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_22/cam_right_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_22/cam_high/frame_000050.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_22/cam_left_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate/episode_22/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl next to the ramekin and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl next to the ramekin and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -54.5, "hop_absolute_value": 54.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate_episode_27-train-bf_000030-af_000010-minus-0000102", "task": "pick up the black bowl from table center and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_35/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_35/cam_high/frame_000134.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_27/cam_high/frame_000030.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_27/cam_left_wrist/frame_000030.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_27/cam_right_wrist/frame_000030.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_27/cam_high/frame_000010.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_27/cam_left_wrist/frame_000010.jpg", "images/libero_data/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate/episode_27/cam_right_wrist/frame_000010.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl from table center and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl from table center and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_cream_cheese_box_and_the_butter_in_the_basket_episode_1-train-bf_000135-af_000025-minus-0000090", "task": "put both the cream cheese box and the butter in the basket", "image": [ "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_44/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_44/cam_high/frame_000275.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_1/cam_high/frame_000135.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_1/cam_left_wrist/frame_000135.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_1/cam_right_wrist/frame_000135.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_1/cam_high/frame_000025.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_1/cam_left_wrist/frame_000025.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_1/cam_right_wrist/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the cream cheese box and the butter in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the cream cheese box and the butter in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -81.5, "hop_absolute_value": 81.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate_episode_25-train-bf_000110-af_000100-minus-0000256", "task": "pick up the black bowl next to the cookie box and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_8/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_8/cam_high/frame_000132.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_25/cam_high/frame_000110.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_25/cam_left_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_25/cam_right_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_25/cam_high/frame_000100.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_25/cam_left_wrist/frame_000100.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_25/cam_right_wrist/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl next to the cookie box and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl next to the cookie box and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate_episode_37-train-bf_000070-af_000025-minus-0000113", "task": "pick up the black bowl on the stove and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_37/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_37/cam_high/frame_000127.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_37/cam_high/frame_000070.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_37/cam_left_wrist/frame_000070.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_37/cam_right_wrist/frame_000070.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_37/cam_high/frame_000025.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_37/cam_left_wrist/frame_000025.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate/episode_37/cam_right_wrist/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the stove and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the stove and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -64.3, "hop_absolute_value": 64.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_turn_on_the_stove_and_put_the_moka_pot_on_it_episode_41-train-bf_000200-af_000155-minus-0000266", "task": "turn on the stove and put the moka pot on it", "image": [ "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_3/cam_high/frame_000000.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_3/cam_high/frame_000284.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_41/cam_high/frame_000200.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_41/cam_left_wrist/frame_000200.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_41/cam_right_wrist/frame_000200.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_41/cam_high/frame_000155.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_41/cam_left_wrist/frame_000155.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_41/cam_right_wrist/frame_000155.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on the stove and put the moka pot on it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on the stove and put the moka pot on it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -22.5, "hop_absolute_value": 22.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_turn_on_the_stove_and_put_the_moka_pot_on_it_episode_36-train-bf_000230-af_000135-minus-0000240", "task": "turn on the stove and put the moka pot on it", "image": [ "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_36/cam_high/frame_000000.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_36/cam_high/frame_000311.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_36/cam_high/frame_000230.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_36/cam_left_wrist/frame_000230.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_36/cam_right_wrist/frame_000230.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_36/cam_high/frame_000135.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_36/cam_left_wrist/frame_000135.jpg", "images/libero_data/turn_on_the_stove_and_put_the_moka_pot_on_it/episode_36/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on the stove and put the moka pot on it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on the stove and put the moka pot on it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -41.3, "hop_absolute_value": 41.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_wine_bottle_on_top_of_the_cabinet_episode_39-train-bf_000085-af_000020-minus-0000058", "task": "put the wine bottle on top of the cabinet", "image": [ "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_39/cam_high/frame_000000.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_39/cam_high/frame_000101.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_39/cam_high/frame_000085.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_39/cam_left_wrist/frame_000085.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_39/cam_right_wrist/frame_000085.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_39/cam_high/frame_000020.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_39/cam_left_wrist/frame_000020.jpg", "images/libero_data/put_the_wine_bottle_on_top_of_the_cabinet/episode_39/cam_right_wrist/frame_000020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the wine bottle on top of the cabinet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the wine bottle on top of the cabinet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -76.5, "hop_absolute_value": 76.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_yellow_and_white_mug_in_the_microwave_and_close_it_episode_46-train-bf_000250-af_000225-minus-0000656", "task": "put the yellow and white mug in the microwave and close it", "image": [ "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_23/cam_high/frame_000000.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_23/cam_high/frame_000342.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_46/cam_high/frame_000250.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_46/cam_left_wrist/frame_000250.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_46/cam_right_wrist/frame_000250.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_46/cam_high/frame_000225.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_46/cam_left_wrist/frame_000225.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_46/cam_right_wrist/frame_000225.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the yellow and white mug in the microwave and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the yellow and white mug in the microwave and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_alphabet_soup_and_place_it_in_the_basket_episode_40-train-bf_000080-af_000065-minus-0000546", "task": "pick up the alphabet soup and place it in the basket", "image": [ "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_21/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_21/cam_high/frame_000159.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_40/cam_high/frame_000080.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_40/cam_left_wrist/frame_000080.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_40/cam_right_wrist/frame_000080.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_40/cam_high/frame_000065.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_40/cam_left_wrist/frame_000065.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_40/cam_right_wrist/frame_000065.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the alphabet soup and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the alphabet soup and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -18.8, "hop_absolute_value": 18.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate_episode_15-train-bf_000192-af_000165-minus-0000715", "task": "pick up the black bowl on the wooden cabinet and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_16/cam_high/frame_000141.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_15/cam_high/frame_000192.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_15/cam_left_wrist/frame_000192.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_15/cam_right_wrist/frame_000192.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_15/cam_high/frame_000165.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_15/cam_left_wrist/frame_000165.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate/episode_15/cam_right_wrist/frame_000165.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the wooden cabinet and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the wooden cabinet and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -13.2, "hop_absolute_value": 13.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_31-train-bf_000160-af_000060-minus-0000151", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_4/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_4/cam_high/frame_000148.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_31/cam_high/frame_000160.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_31/cam_left_wrist/frame_000160.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_31/cam_right_wrist/frame_000160.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_31/cam_high/frame_000060.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_31/cam_left_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_31/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_cream_cheese_and_place_it_in_the_basket_episode_35-train-bf_000115-af_000030-minus-0000110", "task": "pick up the cream cheese and place it in the basket", "image": [ "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_30/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_30/cam_high/frame_000170.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_35/cam_high/frame_000115.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_35/cam_left_wrist/frame_000115.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_35/cam_right_wrist/frame_000115.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_35/cam_high/frame_000030.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_35/cam_left_wrist/frame_000030.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_35/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the cream cheese and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the cream cheese and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.9, "hop_absolute_value": 73.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket_episode_16-train-bf_000105-af_000090-minus-0000274", "task": "put both the alphabet soup and the cream cheese box in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_43/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_43/cam_high/frame_000254.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_16/cam_high/frame_000105.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_16/cam_left_wrist/frame_000105.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_16/cam_right_wrist/frame_000105.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_16/cam_high/frame_000090.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_16/cam_left_wrist/frame_000090.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket/episode_16/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the cream cheese box in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the cream cheese box in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -14.3, "hop_absolute_value": 14.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate_episode_15-train-bf_000190-af_000035-minus-0000095", "task": "put the white mug on the left plate and put the yellow and white mug on the right plate", "image": [ "images/libero_data/put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate/episode_36/cam_high/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate/episode_36/cam_high/frame_000241.jpg", "images/libero_data/put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate/episode_15/cam_high/frame_000190.jpg", "images/libero_data/put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate/episode_15/cam_left_wrist/frame_000190.jpg", "images/libero_data/put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate/episode_15/cam_right_wrist/frame_000190.jpg", "images/libero_data/put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate/episode_15/cam_high/frame_000035.jpg", "images/libero_data/put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate/episode_15/cam_left_wrist/frame_000035.jpg", "images/libero_data/put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate/episode_15/cam_right_wrist/frame_000035.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the white mug on the left plate and put the yellow and white mug on the right plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the white mug on the left plate and put the yellow and white mug on the right plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -81.6, "hop_absolute_value": 81.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it_episode_16-train-bf_000155-af_000120-minus-0000262", "task": "put the black bowl in the bottom drawer of the cabinet and close it", "image": [ "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_27/cam_high/frame_000000.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_27/cam_high/frame_000234.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_16/cam_high/frame_000155.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_16/cam_left_wrist/frame_000155.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_16/cam_right_wrist/frame_000155.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_16/cam_high/frame_000120.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_16/cam_left_wrist/frame_000120.jpg", "images/libero_data/put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it/episode_16/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the black bowl in the bottom drawer of the cabinet and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the black bowl in the bottom drawer of the cabinet and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -22.6, "hop_absolute_value": 22.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_cream_cheese_box_and_the_butter_in_the_basket_episode_10-train-bf_000295-af_000280-minus-0000804", "task": "put both the cream cheese box and the butter in the basket", "image": [ "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_45/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_45/cam_high/frame_000251.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_10/cam_high/frame_000295.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_10/cam_left_wrist/frame_000295.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_10/cam_right_wrist/frame_000295.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_10/cam_high/frame_000280.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_10/cam_left_wrist/frame_000280.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_10/cam_right_wrist/frame_000280.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the cream cheese box and the butter in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the cream cheese box and the butter in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.1, "hop_absolute_value": 5.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate_episode_27-train-bf_000055-af_000030-minus-0000133", "task": "pick up the black bowl between the plate and the ramekin and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_43/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_43/cam_high/frame_000095.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_27/cam_high/frame_000055.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_27/cam_left_wrist/frame_000055.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_27/cam_right_wrist/frame_000055.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_27/cam_high/frame_000030.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_27/cam_left_wrist/frame_000030.jpg", "images/libero_data/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate/episode_27/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl between the plate and the ramekin and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl between the plate and the ramekin and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -45.5, "hop_absolute_value": 45.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_34-train-bf_000164-af_000025-minus-0000062", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_10/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_10/cam_high/frame_000151.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_34/cam_high/frame_000164.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_34/cam_left_wrist/frame_000164.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_34/cam_right_wrist/frame_000164.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_34/cam_high/frame_000025.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_34/cam_left_wrist/frame_000025.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_34/cam_right_wrist/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -84.4, "hop_absolute_value": 84.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_milk_and_place_it_in_the_basket_episode_24-train-bf_000100-af_000025-minus-0000107", "task": "pick up the milk and place it in the basket", "image": [ "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_39/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_39/cam_high/frame_000137.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_24/cam_high/frame_000100.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_24/cam_left_wrist/frame_000100.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_24/cam_right_wrist/frame_000100.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_24/cam_high/frame_000025.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_24/cam_left_wrist/frame_000025.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_24/cam_right_wrist/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the milk and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the milk and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate_episode_37-train-bf_000263-af_000070-minus-0000137", "task": "put the white mug on the plate and put the chocolate pudding to the right of the plate", "image": [ "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_19/cam_high/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_19/cam_high/frame_000214.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_37/cam_high/frame_000263.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_37/cam_left_wrist/frame_000263.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_37/cam_right_wrist/frame_000263.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_37/cam_high/frame_000070.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_37/cam_left_wrist/frame_000070.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_37/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the white mug on the plate and put the chocolate pudding to the right of the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the white mug on the plate and put the chocolate pudding to the right of the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.1, "hop_absolute_value": 73.1, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_butter_and_place_it_in_the_basket_episode_37-train-bf_000110-af_000060-minus-0000803", "task": "pick up the butter and place it in the basket", "image": [ "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_37/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_37/cam_high/frame_000157.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_37/cam_high/frame_000110.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_37/cam_left_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_37/cam_right_wrist/frame_000110.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_37/cam_high/frame_000060.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_37/cam_left_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_butter_and_place_it_in_the_basket/episode_37/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the butter and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the butter and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -45.5, "hop_absolute_value": 45.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_24-train-bf_000035-af_000015-minus-0000161", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_49/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_49/cam_high/frame_000148.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_24/cam_high/frame_000035.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_24/cam_left_wrist/frame_000035.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_24/cam_right_wrist/frame_000035.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_24/cam_high/frame_000015.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_24/cam_left_wrist/frame_000015.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_24/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -57.1, "hop_absolute_value": 57.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_yellow_and_white_mug_in_the_microwave_and_close_it_episode_8-train-bf_000210-af_000030-minus-0000059", "task": "put the yellow and white mug in the microwave and close it", "image": [ "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_37/cam_high/frame_000000.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_37/cam_high/frame_000277.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_8/cam_high/frame_000210.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_8/cam_left_wrist/frame_000210.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_8/cam_right_wrist/frame_000210.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_8/cam_high/frame_000030.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_8/cam_left_wrist/frame_000030.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_8/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the yellow and white mug in the microwave and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the yellow and white mug in the microwave and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_48-train-bf_000130-af_000010-minus-0000035", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_7/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_7/cam_high/frame_000167.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_48/cam_high/frame_000130.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_48/cam_left_wrist/frame_000130.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_48/cam_right_wrist/frame_000130.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_48/cam_high/frame_000010.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_48/cam_left_wrist/frame_000010.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_48/cam_right_wrist/frame_000010.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -92.3, "hop_absolute_value": 92.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_push_the_plate_to_the_front_of_the_stove_episode_0-train-bf_000040-af_000000-minus-0000005", "task": "push the plate to the front of the stove", "image": [ "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_35/cam_high/frame_000000.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_35/cam_high/frame_000168.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_0/cam_high/frame_000040.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_0/cam_left_wrist/frame_000040.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_0/cam_right_wrist/frame_000040.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_0/cam_high/frame_000000.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_0/cam_left_wrist/frame_000000.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_0/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`push the plate to the front of the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `push the plate to the front of the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_orange_juice_and_place_it_in_the_basket_episode_41-train-bf_000080-af_000045-minus-0000173", "task": "pick up the orange juice and place it in the basket", "image": [ "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_3/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_3/cam_high/frame_000141.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_41/cam_high/frame_000080.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_41/cam_left_wrist/frame_000080.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_41/cam_right_wrist/frame_000080.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_41/cam_high/frame_000045.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_41/cam_left_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_41/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the orange juice and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the orange juice and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -43.8, "hop_absolute_value": 43.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate_episode_0-train-bf_000170-af_000120-minus-0000938", "task": "put the white mug on the plate and put the chocolate pudding to the right of the plate", "image": [ "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_0/cam_high/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_0/cam_high/frame_000202.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_0/cam_high/frame_000170.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_0/cam_left_wrist/frame_000170.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_0/cam_right_wrist/frame_000170.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_0/cam_high/frame_000120.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_0/cam_left_wrist/frame_000120.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_0/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the white mug on the plate and put the chocolate pudding to the right of the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the white mug on the plate and put the chocolate pudding to the right of the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -29.4, "hop_absolute_value": 29.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_39-train-bf_000150-af_000115-minus-0000675", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_43/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_43/cam_high/frame_000168.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_39/cam_high/frame_000150.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_39/cam_left_wrist/frame_000150.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_39/cam_right_wrist/frame_000150.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_39/cam_high/frame_000115.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_39/cam_left_wrist/frame_000115.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_39/cam_right_wrist/frame_000115.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -23.3, "hop_absolute_value": 23.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_bbq_sauce_and_place_it_in_the_basket_episode_49-train-bf_000035-af_000030-minus-0000240", "task": "pick up the bbq sauce and place it in the basket", "image": [ "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_48/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_48/cam_high/frame_000169.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_49/cam_high/frame_000035.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_49/cam_left_wrist/frame_000035.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_49/cam_right_wrist/frame_000035.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_49/cam_high/frame_000030.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_49/cam_left_wrist/frame_000030.jpg", "images/libero_data/pick_up_the_bbq_sauce_and_place_it_in_the_basket/episode_49/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the bbq sauce and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the bbq sauce and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -14.3, "hop_absolute_value": 14.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_yellow_and_white_mug_in_the_microwave_and_close_it_episode_24-train-bf_000050-af_000015-minus-0000114", "task": "put the yellow and white mug in the microwave and close it", "image": [ "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_27/cam_high/frame_000000.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_27/cam_high/frame_000372.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_24/cam_high/frame_000050.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_24/cam_left_wrist/frame_000050.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_24/cam_right_wrist/frame_000050.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_24/cam_high/frame_000015.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_24/cam_left_wrist/frame_000015.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_24/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the yellow and white mug in the microwave and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the yellow and white mug in the microwave and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_chocolate_pudding_and_place_it_in_the_basket_episode_6-train-bf_000085-af_000025-minus-0000946", "task": "pick up the chocolate pudding and place it in the basket", "image": [ "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_6/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_6/cam_high/frame_000161.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_6/cam_high/frame_000085.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_6/cam_left_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_6/cam_right_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_6/cam_high/frame_000025.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_6/cam_left_wrist/frame_000025.jpg", "images/libero_data/pick_up_the_chocolate_pudding_and_place_it_in_the_basket/episode_6/cam_right_wrist/frame_000025.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the chocolate pudding and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the chocolate pudding and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -70.6, "hop_absolute_value": 70.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_ketchup_and_place_it_in_the_basket_episode_22-train-bf_000085-af_000000-minus-0000021", "task": "pick up the ketchup and place it in the basket", "image": [ "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_6/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_6/cam_high/frame_000149.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_22/cam_high/frame_000085.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_22/cam_left_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_22/cam_right_wrist/frame_000085.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_22/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_22/cam_left_wrist/frame_000000.jpg", "images/libero_data/pick_up_the_ketchup_and_place_it_in_the_basket/episode_22/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the ketchup and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the ketchup and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_yellow_and_white_mug_in_the_microwave_and_close_it_episode_42-train-bf_000050-af_000035-minus-0000248", "task": "put the yellow and white mug in the microwave and close it", "image": [ "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_49/cam_high/frame_000000.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_49/cam_high/frame_000294.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_42/cam_high/frame_000050.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_42/cam_left_wrist/frame_000050.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_42/cam_right_wrist/frame_000050.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_42/cam_high/frame_000035.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_42/cam_left_wrist/frame_000035.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_42/cam_right_wrist/frame_000035.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the yellow and white mug in the microwave and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the yellow and white mug in the microwave and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -30.0, "hop_absolute_value": 30.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_milk_and_place_it_in_the_basket_episode_14-train-bf_000120-af_000100-minus-0000242", "task": "pick up the milk and place it in the basket", "image": [ "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_2/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_2/cam_high/frame_000142.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_14/cam_high/frame_000120.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_14/cam_left_wrist/frame_000120.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_14/cam_right_wrist/frame_000120.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_14/cam_high/frame_000100.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_14/cam_left_wrist/frame_000100.jpg", "images/libero_data/pick_up_the_milk_and_place_it_in_the_basket/episode_14/cam_right_wrist/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the milk and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the milk and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate_episode_32-train-bf_000035-af_000010-minus-0000055", "task": "pick up the black bowl on the cookie box and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_37/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_37/cam_high/frame_000096.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_32/cam_high/frame_000035.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_32/cam_left_wrist/frame_000035.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_32/cam_right_wrist/frame_000035.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_32/cam_high/frame_000010.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_32/cam_left_wrist/frame_000010.jpg", "images/libero_data/pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate/episode_32/cam_right_wrist/frame_000010.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl on the cookie box and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl on the cookie box and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_salad_dressing_and_place_it_in_the_basket_episode_41-train-bf_000060-af_000015-minus-0000077", "task": "pick up the salad dressing and place it in the basket", "image": [ "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_7/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_7/cam_high/frame_000123.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_41/cam_high/frame_000060.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_41/cam_left_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_41/cam_right_wrist/frame_000060.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_41/cam_high/frame_000015.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_41/cam_left_wrist/frame_000015.jpg", "images/libero_data/pick_up_the_salad_dressing_and_place_it_in_the_basket/episode_41/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the salad dressing and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the salad dressing and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy_episode_34-train-bf_000145-af_000015-minus-0000058", "task": "pick up the book and place it in the back compartment of the caddy", "image": [ "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_34/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_34/cam_high/frame_000177.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_34/cam_high/frame_000145.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_34/cam_left_wrist/frame_000145.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_34/cam_right_wrist/frame_000145.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_34/cam_high/frame_000015.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_34/cam_left_wrist/frame_000015.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_34/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the book and place it in the back compartment of the caddy`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the book and place it in the back compartment of the caddy`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -89.7, "hop_absolute_value": 89.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy_episode_44-train-bf_000130-af_000010-minus-0000020", "task": "pick up the book and place it in the back compartment of the caddy", "image": [ "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_13/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_13/cam_high/frame_000201.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_44/cam_high/frame_000130.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_44/cam_left_wrist/frame_000130.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_44/cam_right_wrist/frame_000130.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_44/cam_high/frame_000010.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_44/cam_left_wrist/frame_000010.jpg", "images/libero_data/pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy/episode_44/cam_right_wrist/frame_000010.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the book and place it in the back compartment of the caddy`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the book and place it in the back compartment of the caddy`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -92.3, "hop_absolute_value": 92.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_cream_cheese_box_and_the_butter_in_the_basket_episode_7-train-bf_000185-af_000030-minus-0000097", "task": "put both the cream cheese box and the butter in the basket", "image": [ "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_7/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_7/cam_high/frame_000245.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_7/cam_high/frame_000185.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_7/cam_left_wrist/frame_000185.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_7/cam_right_wrist/frame_000185.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_7/cam_high/frame_000030.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_7/cam_left_wrist/frame_000030.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_7/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the cream cheese box and the butter in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the cream cheese box and the butter in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.8, "hop_absolute_value": 83.8, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate_episode_11-train-bf_000020-af_000000-minus-0000002", "task": "pick up the black bowl next to the cookie box and place it on the plate", "image": [ "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_11/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_11/cam_high/frame_000110.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_11/cam_high/frame_000020.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_11/cam_left_wrist/frame_000020.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_11/cam_right_wrist/frame_000020.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_11/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_11/cam_left_wrist/frame_000000.jpg", "images/libero_data/pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate/episode_11/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the black bowl next to the cookie box and place it on the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the black bowl next to the cookie box and place it on the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_tomato_sauce_and_place_it_in_the_basket_episode_26-train-bf_000100-af_000040-minus-0000160", "task": "pick up the tomato sauce and place it in the basket", "image": [ "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_26/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_26/cam_high/frame_000160.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_26/cam_high/frame_000100.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_26/cam_left_wrist/frame_000100.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_26/cam_right_wrist/frame_000100.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_26/cam_high/frame_000040.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_26/cam_left_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_26/cam_right_wrist/frame_000040.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the tomato sauce and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the tomato sauce and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_cream_cheese_and_place_it_in_the_basket_episode_10-train-bf_000040-af_000010-minus-0000089", "task": "pick up the cream cheese and place it in the basket", "image": [ "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_40/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_40/cam_high/frame_000149.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_10/cam_high/frame_000040.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_10/cam_left_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_10/cam_right_wrist/frame_000040.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_10/cam_high/frame_000010.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_10/cam_left_wrist/frame_000010.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_10/cam_right_wrist/frame_000010.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the cream cheese and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the cream cheese and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket_episode_49-train-bf_000050-af_000015-minus-0000120", "task": "put both the alphabet soup and the tomato sauce in the basket", "image": [ "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_49/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_49/cam_high/frame_000277.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_49/cam_high/frame_000050.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_49/cam_left_wrist/frame_000050.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_49/cam_right_wrist/frame_000050.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_49/cam_high/frame_000015.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_49/cam_left_wrist/frame_000015.jpg", "images/libero_data/put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket/episode_49/cam_right_wrist/frame_000015.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the alphabet soup and the tomato sauce in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the alphabet soup and the tomato sauce in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_both_the_cream_cheese_box_and_the_butter_in_the_basket_episode_12-train-bf_000245-af_000110-minus-0000200", "task": "put both the cream cheese box and the butter in the basket", "image": [ "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_41/cam_high/frame_000000.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_41/cam_high/frame_000246.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_12/cam_high/frame_000245.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_12/cam_left_wrist/frame_000245.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_12/cam_right_wrist/frame_000245.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_12/cam_high/frame_000110.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_12/cam_left_wrist/frame_000110.jpg", "images/libero_data/put_both_the_cream_cheese_box_and_the_butter_in_the_basket/episode_12/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put both the cream cheese box and the butter in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put both the cream cheese box and the butter in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -55.1, "hop_absolute_value": 55.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "libero_data" }, { "id": "libero_data_push_the_plate_to_the_front_of_the_stove_episode_19-train-bf_000060-af_000050-minus-0000633", "task": "push the plate to the front of the stove", "image": [ "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_27/cam_high/frame_000000.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_27/cam_high/frame_000219.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_19/cam_high/frame_000060.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_19/cam_left_wrist/frame_000060.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_19/cam_right_wrist/frame_000060.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_19/cam_high/frame_000050.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_19/cam_left_wrist/frame_000050.jpg", "images/libero_data/push_the_plate_to_the_front_of_the_stove/episode_19/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`push the plate to the front of the stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `push the plate to the front of the stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_orange_juice_and_place_it_in_the_basket_episode_26-train-bf_000080-af_000075-minus-0000240", "task": "pick up the orange juice and place it in the basket", "image": [ "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_6/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_6/cam_high/frame_000132.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_26/cam_high/frame_000080.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_26/cam_left_wrist/frame_000080.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_26/cam_right_wrist/frame_000080.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_26/cam_high/frame_000075.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_26/cam_left_wrist/frame_000075.jpg", "images/libero_data/pick_up_the_orange_juice_and_place_it_in_the_basket/episode_26/cam_right_wrist/frame_000075.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the orange juice and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the orange juice and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.2, "hop_absolute_value": 6.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_tomato_sauce_and_place_it_in_the_basket_episode_35-train-bf_000050-af_000045-minus-0000521", "task": "pick up the tomato sauce and place it in the basket", "image": [ "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_36/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_36/cam_high/frame_000138.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_35/cam_high/frame_000050.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_35/cam_left_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_35/cam_right_wrist/frame_000050.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_35/cam_high/frame_000045.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_35/cam_left_wrist/frame_000045.jpg", "images/libero_data/pick_up_the_tomato_sauce_and_place_it_in_the_basket/episode_35/cam_right_wrist/frame_000045.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the tomato sauce and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the tomato sauce and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_yellow_and_white_mug_in_the_microwave_and_close_it_episode_40-train-bf_000140-af_000130-minus-0000746", "task": "put the yellow and white mug in the microwave and close it", "image": [ "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_17/cam_high/frame_000000.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_17/cam_high/frame_000278.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_40/cam_high/frame_000140.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_40/cam_left_wrist/frame_000140.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_40/cam_right_wrist/frame_000140.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_40/cam_high/frame_000130.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_40/cam_left_wrist/frame_000130.jpg", "images/libero_data/put_the_yellow_and_white_mug_in_the_microwave_and_close_it/episode_40/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the yellow and white mug in the microwave and close it`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the yellow and white mug in the microwave and close it`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -7.1, "hop_absolute_value": 7.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_alphabet_soup_and_place_it_in_the_basket_episode_25-train-bf_000125-af_000120-minus-0000255", "task": "pick up the alphabet soup and place it in the basket", "image": [ "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_6/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_6/cam_high/frame_000154.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_25/cam_high/frame_000125.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_25/cam_left_wrist/frame_000125.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_25/cam_right_wrist/frame_000125.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_25/cam_high/frame_000120.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_25/cam_left_wrist/frame_000120.jpg", "images/libero_data/pick_up_the_alphabet_soup_and_place_it_in_the_basket/episode_25/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the alphabet soup and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the alphabet soup and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.0, "hop_absolute_value": 4.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_pick_up_the_cream_cheese_and_place_it_in_the_basket_episode_40-train-bf_000140-af_000135-minus-0000267", "task": "pick up the cream cheese and place it in the basket", "image": [ "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_33/cam_high/frame_000000.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_33/cam_high/frame_000143.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_40/cam_high/frame_000140.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_40/cam_left_wrist/frame_000140.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_40/cam_right_wrist/frame_000140.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_40/cam_high/frame_000135.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_40/cam_left_wrist/frame_000135.jpg", "images/libero_data/pick_up_the_cream_cheese_and_place_it_in_the_basket/episode_40/cam_right_wrist/frame_000135.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick up the cream cheese and place it in the basket`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick up the cream cheese and place it in the basket`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.6, "hop_absolute_value": 3.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "libero_data_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate_episode_45-train-bf_000125-af_000120-minus-0000609", "task": "put the white mug on the plate and put the chocolate pudding to the right of the plate", "image": [ "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_24/cam_high/frame_000000.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_24/cam_high/frame_000245.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_45/cam_high/frame_000125.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_45/cam_left_wrist/frame_000125.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_45/cam_right_wrist/frame_000125.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_45/cam_high/frame_000120.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_45/cam_left_wrist/frame_000120.jpg", "images/libero_data/put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate/episode_45/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`put the white mug on the plate and put the chocolate pudding to the right of the plate`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `put the white mug on the plate and put the chocolate pudding to the right of the plate`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.0, "hop_absolute_value": 4.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "libero_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_0-train-bf_000080-af_000280-plus-0000473", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_34/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_34/cam_high/frame_000318.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_high/frame_000080.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_right_wrist/frame_000080.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_high/frame_000280.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_left_wrist/frame_000280.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_right_wrist/frame_000280.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_sink_faucet_episode_11-train-bf_000050-af_000070-plus-0000191", "task": "turn on sink faucet", "image": [ "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_high/frame_000223.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_11/cam_high/frame_000050.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_11/cam_left_wrist/frame_000050.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_11/cam_right_wrist/frame_000050.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_11/cam_high/frame_000070.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_11/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_11/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 14.3, "hop_absolute_value": 14.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_stove_episode_8-train-bf_000110-af_000210-plus-0000909", "task": "turn on stove", "image": [ "images/robocasa_data/turn_on_stove/episode_17/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_stove/episode_17/cam_high/frame_000230.jpg", "images/robocasa_data/turn_on_stove/episode_8/cam_high/frame_000110.jpg", "images/robocasa_data/turn_on_stove/episode_8/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/turn_on_stove/episode_8/cam_right_wrist/frame_000110.jpg", "images/robocasa_data/turn_on_stove/episode_8/cam_high/frame_000210.jpg", "images/robocasa_data/turn_on_stove/episode_8/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/turn_on_stove/episode_8/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 41.7, "hop_absolute_value": 41.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_sink_faucet_episode_19-train-bf_000010-af_000150-plus-0000274", "task": "turn off sink faucet", "image": [ "images/robocasa_data/turn_off_sink_faucet/episode_15/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_15/cam_high/frame_000159.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_19/cam_high/frame_000010.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_19/cam_left_wrist/frame_000010.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_19/cam_right_wrist/frame_000010.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_19/cam_high/frame_000150.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_19/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_19/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_double_door_episode_16-train-bf_000270-af_000350-plus-0000473", "task": "close double door", "image": [ "images/robocasa_data/close_double_door/episode_37/cam_high/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_37/cam_high/frame_000370.jpg", "images/robocasa_data/close_double_door/episode_16/cam_high/frame_000270.jpg", "images/robocasa_data/close_double_door/episode_16/cam_left_wrist/frame_000270.jpg", "images/robocasa_data/close_double_door/episode_16/cam_right_wrist/frame_000270.jpg", "images/robocasa_data/close_double_door/episode_16/cam_high/frame_000350.jpg", "images/robocasa_data/close_double_door/episode_16/cam_left_wrist/frame_000350.jpg", "images/robocasa_data/close_double_door/episode_16/cam_right_wrist/frame_000350.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_single_door_episode_12-train-bf_000110-af_000250-plus-0000407", "task": "open single door", "image": [ "images/robocasa_data/open_single_door/episode_4/cam_high/frame_000000.jpg", "images/robocasa_data/open_single_door/episode_4/cam_high/frame_000309.jpg", "images/robocasa_data/open_single_door/episode_12/cam_high/frame_000110.jpg", "images/robocasa_data/open_single_door/episode_12/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/open_single_door/episode_12/cam_right_wrist/frame_000110.jpg", "images/robocasa_data/open_single_door/episode_12/cam_high/frame_000250.jpg", "images/robocasa_data/open_single_door/episode_12/cam_left_wrist/frame_000250.jpg", "images/robocasa_data/open_single_door/episode_12/cam_right_wrist/frame_000250.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_stove_to_counter_episode_5-train-bf_000020-af_000080-plus-0000282", "task": "pick and place stove to counter", "image": [ "images/robocasa_data/pick_and_place_stove_to_counter/episode_18/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_18/cam_high/frame_000289.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_5/cam_high/frame_000020.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_5/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_5/cam_right_wrist/frame_000020.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_5/cam_high/frame_000080.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_5/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_5/cam_right_wrist/frame_000080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place stove to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place stove to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 23.1, "hop_absolute_value": 23.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_microwave_to_counter_episode_15-train-bf_000220-af_000290-plus-0000384", "task": "pick and place microwave to counter", "image": [ "images/robocasa_data/pick_and_place_microwave_to_counter/episode_7/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_7/cam_high/frame_000324.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_high/frame_000220.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_left_wrist/frame_000220.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_right_wrist/frame_000220.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_high/frame_000290.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_left_wrist/frame_000290.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_right_wrist/frame_000290.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place microwave to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place microwave to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 53.8, "hop_absolute_value": 53.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_stove_episode_35-train-bf_000170-af_000220-plus-0000329", "task": "pick and place counter to stove", "image": [ "images/robocasa_data/pick_and_place_counter_to_stove/episode_23/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_23/cam_high/frame_000316.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_35/cam_high/frame_000170.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_35/cam_left_wrist/frame_000170.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_35/cam_right_wrist/frame_000170.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_35/cam_high/frame_000220.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_35/cam_left_wrist/frame_000220.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_35/cam_right_wrist/frame_000220.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 41.7, "hop_absolute_value": 41.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_stove_to_counter_episode_39-train-bf_000010-af_000090-plus-0000276", "task": "pick and place stove to counter", "image": [ "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_high/frame_000385.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_39/cam_high/frame_000010.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_39/cam_left_wrist/frame_000010.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_39/cam_right_wrist/frame_000010.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_39/cam_high/frame_000090.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_39/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_39/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place stove to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place stove to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 36.4, "hop_absolute_value": 36.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_16-train-bf_000130-af_000480-plus-0000482", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_0/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_0/cam_high/frame_000730.jpg", "images/robocasa_data/open_double_door/episode_16/cam_high/frame_000130.jpg", "images/robocasa_data/open_double_door/episode_16/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/open_double_door/episode_16/cam_right_wrist/frame_000130.jpg", "images/robocasa_data/open_double_door/episode_16/cam_high/frame_000480.jpg", "images/robocasa_data/open_double_door/episode_16/cam_left_wrist/frame_000480.jpg", "images/robocasa_data/open_double_door/episode_16/cam_right_wrist/frame_000480.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_sink_faucet_episode_10-train-bf_000180-af_000223-plus-0000391", "task": "turn on sink faucet", "image": [ "images/robocasa_data/turn_on_sink_faucet/episode_2/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_2/cam_high/frame_000266.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_10/cam_high/frame_000180.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_10/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_10/cam_right_wrist/frame_000180.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_10/cam_high/frame_000223.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_10/cam_left_wrist/frame_000223.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_10/cam_right_wrist/frame_000223.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_double_door_episode_9-train-bf_000180-af_000220-plus-0000754", "task": "close double door", "image": [ "images/robocasa_data/close_double_door/episode_26/cam_high/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_26/cam_high/frame_000417.jpg", "images/robocasa_data/close_double_door/episode_9/cam_high/frame_000180.jpg", "images/robocasa_data/close_double_door/episode_9/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/close_double_door/episode_9/cam_right_wrist/frame_000180.jpg", "images/robocasa_data/close_double_door/episode_9/cam_high/frame_000220.jpg", "images/robocasa_data/close_double_door/episode_9/cam_left_wrist/frame_000220.jpg", "images/robocasa_data/close_double_door/episode_9/cam_right_wrist/frame_000220.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 19.0, "hop_absolute_value": 19.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_setup_mug_episode_7-train-bf_000250-af_000310-plus-0000385", "task": "coffee setup mug", "image": [ "images/robocasa_data/coffee_setup_mug/episode_26/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_26/cam_high/frame_000229.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_high/frame_000250.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_left_wrist/frame_000250.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_right_wrist/frame_000250.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_high/frame_000310.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_left_wrist/frame_000310.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_right_wrist/frame_000310.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee setup mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee setup mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 54.5, "hop_absolute_value": 54.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_double_door_episode_4-train-bf_000000-af_000190-plus-0000401", "task": "close double door", "image": [ "images/robocasa_data/close_double_door/episode_40/cam_high/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_40/cam_high/frame_000383.jpg", "images/robocasa_data/close_double_door/episode_4/cam_high/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_4/cam_left_wrist/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_4/cam_right_wrist/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_4/cam_high/frame_000190.jpg", "images/robocasa_data/close_double_door/episode_4/cam_left_wrist/frame_000190.jpg", "images/robocasa_data/close_double_door/episode_4/cam_right_wrist/frame_000190.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 55.9, "hop_absolute_value": 55.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_microwave_episode_4-train-bf_000020-af_000110-plus-0000315", "task": "turn off microwave", "image": [ "images/robocasa_data/turn_off_microwave/episode_4/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_high/frame_000297.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_high/frame_000020.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_right_wrist/frame_000020.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_high/frame_000110.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_sink_spout_episode_3-train-bf_000040-af_000140-plus-0000235", "task": "turn sink spout", "image": [ "images/robocasa_data/turn_sink_spout/episode_20/cam_high/frame_000000.jpg", "images/robocasa_data/turn_sink_spout/episode_20/cam_high/frame_000157.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_high/frame_000040.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_right_wrist/frame_000040.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_high/frame_000140.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_left_wrist/frame_000140.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_right_wrist/frame_000140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn sink spout`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn sink spout`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_stove_to_counter_episode_2-train-bf_000050-af_000190-plus-0000396", "task": "pick and place stove to counter", "image": [ "images/robocasa_data/pick_and_place_stove_to_counter/episode_2/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_2/cam_high/frame_000250.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_2/cam_high/frame_000050.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_2/cam_left_wrist/frame_000050.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_2/cam_right_wrist/frame_000050.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_2/cam_high/frame_000190.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_2/cam_left_wrist/frame_000190.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_2/cam_right_wrist/frame_000190.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place stove to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place stove to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_28-train-bf_000270-af_000290-plus-0000824", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_30/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_30/cam_high/frame_000373.jpg", "images/robocasa_data/coffee_serve_mug/episode_28/cam_high/frame_000270.jpg", "images/robocasa_data/coffee_serve_mug/episode_28/cam_left_wrist/frame_000270.jpg", "images/robocasa_data/coffee_serve_mug/episode_28/cam_right_wrist/frame_000270.jpg", "images/robocasa_data/coffee_serve_mug/episode_28/cam_high/frame_000290.jpg", "images/robocasa_data/coffee_serve_mug/episode_28/cam_left_wrist/frame_000290.jpg", "images/robocasa_data/coffee_serve_mug/episode_28/cam_right_wrist/frame_000290.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 22.2, "hop_absolute_value": 22.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_microwave_to_counter_episode_3-train-bf_000130-af_000230-plus-0000373", "task": "pick and place microwave to counter", "image": [ "images/robocasa_data/pick_and_place_microwave_to_counter/episode_37/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_37/cam_high/frame_000368.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_3/cam_high/frame_000130.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_3/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_3/cam_right_wrist/frame_000130.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_3/cam_high/frame_000230.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_3/cam_left_wrist/frame_000230.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_3/cam_right_wrist/frame_000230.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place microwave to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place microwave to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_12-train-bf_000110-af_000230-plus-0000423", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_12/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_12/cam_high/frame_000259.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_12/cam_high/frame_000110.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_12/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_12/cam_right_wrist/frame_000110.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_12/cam_high/frame_000230.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_12/cam_left_wrist/frame_000230.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_12/cam_right_wrist/frame_000230.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_drawer_episode_387-train-bf_000120-af_000190-plus-0000337", "task": "close drawer", "image": [ "images/robocasa_data/close_drawer/episode_399/cam_high/frame_000000.jpg", "images/robocasa_data/close_drawer/episode_399/cam_high/frame_000158.jpg", "images/robocasa_data/close_drawer/episode_387/cam_high/frame_000120.jpg", "images/robocasa_data/close_drawer/episode_387/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/close_drawer/episode_387/cam_right_wrist/frame_000120.jpg", "images/robocasa_data/close_drawer/episode_387/cam_high/frame_000190.jpg", "images/robocasa_data/close_drawer/episode_387/cam_left_wrist/frame_000190.jpg", "images/robocasa_data/close_drawer/episode_387/cam_right_wrist/frame_000190.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_setup_mug_episode_27-train-bf_000000-af_000130-plus-0000796", "task": "coffee setup mug", "image": [ "images/robocasa_data/coffee_setup_mug/episode_61/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_61/cam_high/frame_000275.jpg", "images/robocasa_data/coffee_setup_mug/episode_27/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_27/cam_left_wrist/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_27/cam_right_wrist/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_27/cam_high/frame_000130.jpg", "images/robocasa_data/coffee_setup_mug/episode_27/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/coffee_setup_mug/episode_27/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee setup mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee setup mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 41.9, "hop_absolute_value": 41.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_setup_mug_episode_24-train-bf_000040-af_000190-plus-0000371", "task": "coffee setup mug", "image": [ "images/robocasa_data/coffee_setup_mug/episode_50/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_50/cam_high/frame_000339.jpg", "images/robocasa_data/coffee_setup_mug/episode_24/cam_high/frame_000040.jpg", "images/robocasa_data/coffee_setup_mug/episode_24/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/coffee_setup_mug/episode_24/cam_right_wrist/frame_000040.jpg", "images/robocasa_data/coffee_setup_mug/episode_24/cam_high/frame_000190.jpg", "images/robocasa_data/coffee_setup_mug/episode_24/cam_left_wrist/frame_000190.jpg", "images/robocasa_data/coffee_setup_mug/episode_24/cam_right_wrist/frame_000190.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee setup mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee setup mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 51.7, "hop_absolute_value": 51.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_single_door_episode_23-train-bf_000020-af_000030-plus-0000189", "task": "close single door", "image": [ "images/robocasa_data/close_single_door/episode_4/cam_high/frame_000000.jpg", "images/robocasa_data/close_single_door/episode_4/cam_high/frame_000261.jpg", "images/robocasa_data/close_single_door/episode_23/cam_high/frame_000020.jpg", "images/robocasa_data/close_single_door/episode_23/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/close_single_door/episode_23/cam_right_wrist/frame_000020.jpg", "images/robocasa_data/close_single_door/episode_23/cam_high/frame_000030.jpg", "images/robocasa_data/close_single_door/episode_23/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/close_single_door/episode_23/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 5.6, "hop_absolute_value": 5.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_sink_to_counter_episode_17-train-bf_000100-af_000180-plus-0000582", "task": "pick and place sink to counter", "image": [ "images/robocasa_data/pick_and_place_sink_to_counter/episode_32/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_32/cam_high/frame_000260.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_17/cam_high/frame_000100.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_17/cam_left_wrist/frame_000100.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_17/cam_right_wrist/frame_000100.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_17/cam_high/frame_000180.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_17/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_17/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place sink to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place sink to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 57.1, "hop_absolute_value": 57.1, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_microwave_episode_17-train-bf_000020-af_000120-plus-0000661", "task": "turn on microwave", "image": [ "images/robocasa_data/turn_on_microwave/episode_27/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_microwave/episode_27/cam_high/frame_000242.jpg", "images/robocasa_data/turn_on_microwave/episode_17/cam_high/frame_000020.jpg", "images/robocasa_data/turn_on_microwave/episode_17/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/turn_on_microwave/episode_17/cam_right_wrist/frame_000020.jpg", "images/robocasa_data/turn_on_microwave/episode_17/cam_high/frame_000120.jpg", "images/robocasa_data/turn_on_microwave/episode_17/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/turn_on_microwave/episode_17/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_drawer_episode_2461-train-bf_000000-af_000020-plus-0000352", "task": "open drawer", "image": [ "images/robocasa_data/open_drawer/episode_2457/cam_high/frame_000000.jpg", "images/robocasa_data/open_drawer/episode_2457/cam_high/frame_000246.jpg", "images/robocasa_data/open_drawer/episode_2461/cam_high/frame_000000.jpg", "images/robocasa_data/open_drawer/episode_2461/cam_left_wrist/frame_000000.jpg", "images/robocasa_data/open_drawer/episode_2461/cam_right_wrist/frame_000000.jpg", "images/robocasa_data/open_drawer/episode_2461/cam_high/frame_000020.jpg", "images/robocasa_data/open_drawer/episode_2461/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/open_drawer/episode_2461/cam_right_wrist/frame_000020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.5, "hop_absolute_value": 10.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_30-train-bf_000560-af_000680-plus-0000496", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_9/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_9/cam_high/frame_000698.jpg", "images/robocasa_data/open_double_door/episode_30/cam_high/frame_000560.jpg", "images/robocasa_data/open_double_door/episode_30/cam_left_wrist/frame_000560.jpg", "images/robocasa_data/open_double_door/episode_30/cam_right_wrist/frame_000560.jpg", "images/robocasa_data/open_double_door/episode_30/cam_high/frame_000680.jpg", "images/robocasa_data/open_double_door/episode_30/cam_left_wrist/frame_000680.jpg", "images/robocasa_data/open_double_door/episode_30/cam_right_wrist/frame_000680.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_microwave_episode_2-train-bf_000140-af_000360-plus-0000462", "task": "pick and place counter to microwave", "image": [ "images/robocasa_data/pick_and_place_counter_to_microwave/episode_46/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_46/cam_high/frame_000430.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_2/cam_high/frame_000140.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_2/cam_left_wrist/frame_000140.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_2/cam_right_wrist/frame_000140.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_2/cam_high/frame_000360.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_2/cam_left_wrist/frame_000360.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_2/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 68.8, "hop_absolute_value": 68.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_20-train-bf_000020-af_000090-plus-0000309", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_13/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_13/cam_high/frame_000343.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_20/cam_high/frame_000020.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_20/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_20/cam_right_wrist/frame_000020.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_20/cam_high/frame_000090.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_20/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_20/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 17.5, "hop_absolute_value": 17.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_microwave_episode_29-train-bf_000030-af_000070-plus-0000185", "task": "turn on microwave", "image": [ "images/robocasa_data/turn_on_microwave/episode_14/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_microwave/episode_14/cam_high/frame_000216.jpg", "images/robocasa_data/turn_on_microwave/episode_29/cam_high/frame_000030.jpg", "images/robocasa_data/turn_on_microwave/episode_29/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/turn_on_microwave/episode_29/cam_right_wrist/frame_000030.jpg", "images/robocasa_data/turn_on_microwave/episode_29/cam_high/frame_000070.jpg", "images/robocasa_data/turn_on_microwave/episode_29/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/turn_on_microwave/episode_29/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 26.7, "hop_absolute_value": 26.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_sink_episode_43-train-bf_000050-af_000120-plus-0000812", "task": "pick and place counter to sink", "image": [ "images/robocasa_data/pick_and_place_counter_to_sink/episode_3/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_3/cam_high/frame_000498.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_43/cam_high/frame_000050.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_43/cam_left_wrist/frame_000050.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_43/cam_right_wrist/frame_000050.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_43/cam_high/frame_000120.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_43/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_43/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 18.9, "hop_absolute_value": 18.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_9-train-bf_000040-af_000390-plus-0000433", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_6/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_6/cam_high/frame_000727.jpg", "images/robocasa_data/open_double_door/episode_9/cam_high/frame_000040.jpg", "images/robocasa_data/open_double_door/episode_9/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/open_double_door/episode_9/cam_right_wrist/frame_000040.jpg", "images/robocasa_data/open_double_door/episode_9/cam_high/frame_000390.jpg", "images/robocasa_data/open_double_door/episode_9/cam_left_wrist/frame_000390.jpg", "images/robocasa_data/open_double_door/episode_9/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 53.8, "hop_absolute_value": 53.8, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_microwave_episode_30-train-bf_000250-af_000260-plus-0000463", "task": "turn on microwave", "image": [ "images/robocasa_data/turn_on_microwave/episode_16/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_microwave/episode_16/cam_high/frame_000186.jpg", "images/robocasa_data/turn_on_microwave/episode_30/cam_high/frame_000250.jpg", "images/robocasa_data/turn_on_microwave/episode_30/cam_left_wrist/frame_000250.jpg", "images/robocasa_data/turn_on_microwave/episode_30/cam_right_wrist/frame_000250.jpg", "images/robocasa_data/turn_on_microwave/episode_30/cam_high/frame_000260.jpg", "images/robocasa_data/turn_on_microwave/episode_30/cam_left_wrist/frame_000260.jpg", "images/robocasa_data/turn_on_microwave/episode_30/cam_right_wrist/frame_000260.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_cab_episode_18-train-bf_000130-af_000180-plus-0000805", "task": "pick and place counter to cab", "image": [ "images/robocasa_data/pick_and_place_counter_to_cab/episode_6/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_6/cam_high/frame_000242.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_18/cam_high/frame_000130.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_18/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_18/cam_right_wrist/frame_000130.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_18/cam_high/frame_000180.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_18/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_18/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to cab`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to cab`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 22.7, "hop_absolute_value": 22.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_stove_to_counter_episode_16-train-bf_000070-af_000340-plus-0000524", "task": "pick and place stove to counter", "image": [ "images/robocasa_data/pick_and_place_stove_to_counter/episode_7/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_7/cam_high/frame_000264.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_high/frame_000070.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_right_wrist/frame_000070.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_high/frame_000340.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_left_wrist/frame_000340.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_right_wrist/frame_000340.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place stove to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place stove to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 87.1, "hop_absolute_value": 87.1, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_double_door_episode_37-train-bf_000030-af_000130-plus-0000337", "task": "close double door", "image": [ "images/robocasa_data/close_double_door/episode_12/cam_high/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_12/cam_high/frame_000365.jpg", "images/robocasa_data/close_double_door/episode_37/cam_high/frame_000030.jpg", "images/robocasa_data/close_double_door/episode_37/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/close_double_door/episode_37/cam_right_wrist/frame_000030.jpg", "images/robocasa_data/close_double_door/episode_37/cam_high/frame_000130.jpg", "images/robocasa_data/close_double_door/episode_37/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/close_double_door/episode_37/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 29.4, "hop_absolute_value": 29.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_sink_faucet_episode_1-train-bf_000010-af_000110-plus-0000302", "task": "turn off sink faucet", "image": [ "images/robocasa_data/turn_off_sink_faucet/episode_22/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_22/cam_high/frame_000326.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_1/cam_high/frame_000010.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_1/cam_left_wrist/frame_000010.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_1/cam_right_wrist/frame_000010.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_1/cam_high/frame_000110.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_1/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_1/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 45.5, "hop_absolute_value": 45.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_cab_episode_19-train-bf_000180-af_000240-plus-0000789", "task": "pick and place counter to cab", "image": [ "images/robocasa_data/pick_and_place_counter_to_cab/episode_19/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_19/cam_high/frame_000301.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_19/cam_high/frame_000180.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_19/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_19/cam_right_wrist/frame_000180.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_19/cam_high/frame_000240.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_19/cam_left_wrist/frame_000240.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_19/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to cab`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to cab`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_double_door_episode_40-train-bf_000150-af_000180-plus-0000764", "task": "close double door", "image": [ "images/robocasa_data/close_double_door/episode_40/cam_high/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_40/cam_high/frame_000383.jpg", "images/robocasa_data/close_double_door/episode_40/cam_high/frame_000150.jpg", "images/robocasa_data/close_double_door/episode_40/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/close_double_door/episode_40/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/close_double_door/episode_40/cam_high/frame_000180.jpg", "images/robocasa_data/close_double_door/episode_40/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/close_double_door/episode_40/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 13.0, "hop_absolute_value": 13.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_microwave_to_counter_episode_41-train-bf_000080-af_000160-plus-0000315", "task": "pick and place microwave to counter", "image": [ "images/robocasa_data/pick_and_place_microwave_to_counter/episode_19/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_19/cam_high/frame_000327.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_41/cam_high/frame_000080.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_41/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_41/cam_right_wrist/frame_000080.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_41/cam_high/frame_000160.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_41/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_41/cam_right_wrist/frame_000160.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place microwave to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place microwave to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_22-train-bf_000140-af_000170-plus-0000658", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_22/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_22/cam_high/frame_000319.jpg", "images/robocasa_data/coffee_serve_mug/episode_22/cam_high/frame_000140.jpg", "images/robocasa_data/coffee_serve_mug/episode_22/cam_left_wrist/frame_000140.jpg", "images/robocasa_data/coffee_serve_mug/episode_22/cam_right_wrist/frame_000140.jpg", "images/robocasa_data/coffee_serve_mug/episode_22/cam_high/frame_000170.jpg", "images/robocasa_data/coffee_serve_mug/episode_22/cam_left_wrist/frame_000170.jpg", "images/robocasa_data/coffee_serve_mug/episode_22/cam_right_wrist/frame_000170.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 17.6, "hop_absolute_value": 17.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_sink_episode_33-train-bf_000000-af_000080-plus-0000316", "task": "pick and place counter to sink", "image": [ "images/robocasa_data/pick_and_place_counter_to_sink/episode_33/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_33/cam_high/frame_000506.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_33/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_33/cam_left_wrist/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_33/cam_right_wrist/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_33/cam_high/frame_000080.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_33/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_33/cam_right_wrist/frame_000080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.0, "hop_absolute_value": 16.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_5-train-bf_000060-af_000110-plus-0000298", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_3/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_3/cam_high/frame_000296.jpg", "images/robocasa_data/coffee_serve_mug/episode_5/cam_high/frame_000060.jpg", "images/robocasa_data/coffee_serve_mug/episode_5/cam_left_wrist/frame_000060.jpg", "images/robocasa_data/coffee_serve_mug/episode_5/cam_right_wrist/frame_000060.jpg", "images/robocasa_data/coffee_serve_mug/episode_5/cam_high/frame_000110.jpg", "images/robocasa_data/coffee_serve_mug/episode_5/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/coffee_serve_mug/episode_5/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 17.2, "hop_absolute_value": 17.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_setup_mug_episode_61-train-bf_000100-af_000190-plus-0000701", "task": "coffee setup mug", "image": [ "images/robocasa_data/coffee_setup_mug/episode_61/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_61/cam_high/frame_000275.jpg", "images/robocasa_data/coffee_setup_mug/episode_61/cam_high/frame_000100.jpg", "images/robocasa_data/coffee_setup_mug/episode_61/cam_left_wrist/frame_000100.jpg", "images/robocasa_data/coffee_setup_mug/episode_61/cam_right_wrist/frame_000100.jpg", "images/robocasa_data/coffee_setup_mug/episode_61/cam_high/frame_000190.jpg", "images/robocasa_data/coffee_setup_mug/episode_61/cam_left_wrist/frame_000190.jpg", "images/robocasa_data/coffee_setup_mug/episode_61/cam_right_wrist/frame_000190.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee setup mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee setup mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 52.9, "hop_absolute_value": 52.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_microwave_episode_2-train-bf_000050-af_000160-plus-0000325", "task": "pick and place counter to microwave", "image": [ "images/robocasa_data/pick_and_place_counter_to_microwave/episode_14/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_14/cam_high/frame_000429.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_2/cam_high/frame_000050.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_2/cam_left_wrist/frame_000050.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_2/cam_right_wrist/frame_000050.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_2/cam_high/frame_000160.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_2/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_2/cam_right_wrist/frame_000160.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 26.8, "hop_absolute_value": 26.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_stove_to_counter_episode_20-train-bf_000000-af_000060-plus-0000673", "task": "pick and place stove to counter", "image": [ "images/robocasa_data/pick_and_place_stove_to_counter/episode_21/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_21/cam_high/frame_000290.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_left_wrist/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_right_wrist/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_high/frame_000060.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_left_wrist/frame_000060.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place stove to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place stove to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 18.2, "hop_absolute_value": 18.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_single_door_episode_28-train-bf_000150-af_000170-plus-0000607", "task": "close single door", "image": [ "images/robocasa_data/close_single_door/episode_25/cam_high/frame_000000.jpg", "images/robocasa_data/close_single_door/episode_25/cam_high/frame_000283.jpg", "images/robocasa_data/close_single_door/episode_28/cam_high/frame_000150.jpg", "images/robocasa_data/close_single_door/episode_28/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/close_single_door/episode_28/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/close_single_door/episode_28/cam_high/frame_000170.jpg", "images/robocasa_data/close_single_door/episode_28/cam_left_wrist/frame_000170.jpg", "images/robocasa_data/close_single_door/episode_28/cam_right_wrist/frame_000170.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.8, "hop_absolute_value": 11.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_single_door_episode_26-train-bf_000030-af_000110-plus-0000264", "task": "close single door", "image": [ "images/robocasa_data/close_single_door/episode_7/cam_high/frame_000000.jpg", "images/robocasa_data/close_single_door/episode_7/cam_high/frame_000311.jpg", "images/robocasa_data/close_single_door/episode_26/cam_high/frame_000030.jpg", "images/robocasa_data/close_single_door/episode_26/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/close_single_door/episode_26/cam_right_wrist/frame_000030.jpg", "images/robocasa_data/close_single_door/episode_26/cam_high/frame_000110.jpg", "images/robocasa_data/close_single_door/episode_26/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/close_single_door/episode_26/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 42.1, "hop_absolute_value": 42.1, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_microwave_episode_22-train-bf_000130-af_000440-plus-0000573", "task": "pick and place counter to microwave", "image": [ "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_high/frame_000461.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_high/frame_000130.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_right_wrist/frame_000130.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_high/frame_000440.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_left_wrist/frame_000440.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_right_wrist/frame_000440.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 93.9, "hop_absolute_value": 93.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_cab_episode_33-train-bf_000020-af_000160-plus-0000340", "task": "pick and place counter to cab", "image": [ "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_high/frame_000224.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_high/frame_000020.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_right_wrist/frame_000020.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_high/frame_000160.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_right_wrist/frame_000160.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to cab`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to cab`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_sink_faucet_episode_2-train-bf_000100-af_000180-plus-0000649", "task": "turn on sink faucet", "image": [ "images/robocasa_data/turn_on_sink_faucet/episode_23/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_23/cam_high/frame_000226.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_2/cam_high/frame_000100.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_2/cam_left_wrist/frame_000100.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_2/cam_right_wrist/frame_000100.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_2/cam_high/frame_000180.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_2/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_2/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_sink_to_counter_episode_10-train-bf_000120-af_000210-plus-0000383", "task": "pick and place sink to counter", "image": [ "images/robocasa_data/pick_and_place_sink_to_counter/episode_0/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_0/cam_high/frame_000245.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_10/cam_high/frame_000120.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_10/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_10/cam_right_wrist/frame_000120.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_10/cam_high/frame_000210.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_10/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_10/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place sink to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place sink to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 69.2, "hop_absolute_value": 69.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_microwave_to_counter_episode_10-train-bf_000010-af_000270-plus-0000481", "task": "pick and place microwave to counter", "image": [ "images/robocasa_data/pick_and_place_microwave_to_counter/episode_10/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_10/cam_high/frame_000333.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_10/cam_high/frame_000010.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_10/cam_left_wrist/frame_000010.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_10/cam_right_wrist/frame_000010.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_10/cam_high/frame_000270.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_10/cam_left_wrist/frame_000270.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_10/cam_right_wrist/frame_000270.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place microwave to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place microwave to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 81.2, "hop_absolute_value": 81.2, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_sink_to_counter_episode_25-train-bf_000020-af_000210-plus-0000418", "task": "pick and place sink to counter", "image": [ "images/robocasa_data/pick_and_place_sink_to_counter/episode_10/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_10/cam_high/frame_000250.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_25/cam_high/frame_000020.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_25/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_25/cam_right_wrist/frame_000020.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_25/cam_high/frame_000210.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_25/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_25/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place sink to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place sink to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 86.4, "hop_absolute_value": 86.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_single_door_episode_0-train-bf_000160-af_000180-plus-0000535", "task": "open single door", "image": [ "images/robocasa_data/open_single_door/episode_22/cam_high/frame_000000.jpg", "images/robocasa_data/open_single_door/episode_22/cam_high/frame_000213.jpg", "images/robocasa_data/open_single_door/episode_0/cam_high/frame_000160.jpg", "images/robocasa_data/open_single_door/episode_0/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/open_single_door/episode_0/cam_right_wrist/frame_000160.jpg", "images/robocasa_data/open_single_door/episode_0/cam_high/frame_000180.jpg", "images/robocasa_data/open_single_door/episode_0/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/open_single_door/episode_0/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_single_door_episode_11-train-bf_000080-af_000150-plus-0000293", "task": "close single door", "image": [ "images/robocasa_data/close_single_door/episode_6/cam_high/frame_000000.jpg", "images/robocasa_data/close_single_door/episode_6/cam_high/frame_000284.jpg", "images/robocasa_data/close_single_door/episode_11/cam_high/frame_000080.jpg", "images/robocasa_data/close_single_door/episode_11/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/close_single_door/episode_11/cam_right_wrist/frame_000080.jpg", "images/robocasa_data/close_single_door/episode_11/cam_high/frame_000150.jpg", "images/robocasa_data/close_single_door/episode_11/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/close_single_door/episode_11/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 46.7, "hop_absolute_value": 46.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_microwave_to_counter_episode_20-train-bf_000010-af_000220-plus-0000912", "task": "pick and place microwave to counter", "image": [ "images/robocasa_data/pick_and_place_microwave_to_counter/episode_7/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_7/cam_high/frame_000324.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_20/cam_high/frame_000010.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_20/cam_left_wrist/frame_000010.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_20/cam_right_wrist/frame_000010.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_20/cam_high/frame_000220.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_20/cam_left_wrist/frame_000220.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_20/cam_right_wrist/frame_000220.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place microwave to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place microwave to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.0, "hop_absolute_value": 70.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_stove_episode_12-train-bf_000120-af_000240-plus-0000400", "task": "pick and place counter to stove", "image": [ "images/robocasa_data/pick_and_place_counter_to_stove/episode_6/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_6/cam_high/frame_000256.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_12/cam_high/frame_000120.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_12/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_12/cam_right_wrist/frame_000120.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_12/cam_high/frame_000240.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_12/cam_left_wrist/frame_000240.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_12/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_41-train-bf_000090-af_000100-plus-0000628", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_32/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_32/cam_high/frame_000749.jpg", "images/robocasa_data/open_double_door/episode_41/cam_high/frame_000090.jpg", "images/robocasa_data/open_double_door/episode_41/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/open_double_door/episode_41/cam_right_wrist/frame_000090.jpg", "images/robocasa_data/open_double_door/episode_41/cam_high/frame_000100.jpg", "images/robocasa_data/open_double_door/episode_41/cam_left_wrist/frame_000100.jpg", "images/robocasa_data/open_double_door/episode_41/cam_right_wrist/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 1.7, "hop_absolute_value": 1.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_5-train-bf_000060-af_000240-plus-0000457", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_5/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_5/cam_high/frame_000299.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_5/cam_high/frame_000060.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_5/cam_left_wrist/frame_000060.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_5/cam_right_wrist/frame_000060.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_5/cam_high/frame_000240.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_5/cam_left_wrist/frame_000240.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_5/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 78.3, "hop_absolute_value": 78.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_30-train-bf_000180-af_000210-plus-0000693", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_11/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_11/cam_high/frame_000305.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_30/cam_high/frame_000180.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_30/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_30/cam_right_wrist/frame_000180.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_30/cam_high/frame_000210.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_30/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_30/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 23.1, "hop_absolute_value": 23.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_single_door_episode_22-train-bf_000090-af_000180-plus-0000321", "task": "open single door", "image": [ "images/robocasa_data/open_single_door/episode_14/cam_high/frame_000000.jpg", "images/robocasa_data/open_single_door/episode_14/cam_high/frame_000337.jpg", "images/robocasa_data/open_single_door/episode_22/cam_high/frame_000090.jpg", "images/robocasa_data/open_single_door/episode_22/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/open_single_door/episode_22/cam_right_wrist/frame_000090.jpg", "images/robocasa_data/open_single_door/episode_22/cam_high/frame_000180.jpg", "images/robocasa_data/open_single_door/episode_22/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/open_single_door/episode_22/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_press_button_episode_28-train-bf_000050-af_000060-plus-0000070", "task": "coffee press button", "image": [ "images/robocasa_data/coffee_press_button/episode_19/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_press_button/episode_19/cam_high/frame_000125.jpg", "images/robocasa_data/coffee_press_button/episode_28/cam_high/frame_000050.jpg", "images/robocasa_data/coffee_press_button/episode_28/cam_left_wrist/frame_000050.jpg", "images/robocasa_data/coffee_press_button/episode_28/cam_right_wrist/frame_000050.jpg", "images/robocasa_data/coffee_press_button/episode_28/cam_high/frame_000060.jpg", "images/robocasa_data/coffee_press_button/episode_28/cam_left_wrist/frame_000060.jpg", "images/robocasa_data/coffee_press_button/episode_28/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee press button`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee press button`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_30-train-bf_000090-af_000250-plus-0000442", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_18/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_18/cam_high/frame_000334.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_30/cam_high/frame_000090.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_30/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_30/cam_right_wrist/frame_000090.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_30/cam_high/frame_000250.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_30/cam_left_wrist/frame_000250.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_30/cam_right_wrist/frame_000250.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 72.7, "hop_absolute_value": 72.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_stove_episode_2-train-bf_000090-af_000160-plus-0000840", "task": "pick and place counter to stove", "image": [ "images/robocasa_data/pick_and_place_counter_to_stove/episode_12/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_12/cam_high/frame_000308.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_2/cam_high/frame_000090.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_2/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_2/cam_right_wrist/frame_000090.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_2/cam_high/frame_000160.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_2/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_2/cam_right_wrist/frame_000160.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 28.0, "hop_absolute_value": 28.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_stove_episode_35-train-bf_000070-af_000160-plus-0000341", "task": "turn on stove", "image": [ "images/robocasa_data/turn_on_stove/episode_39/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_stove/episode_39/cam_high/frame_000189.jpg", "images/robocasa_data/turn_on_stove/episode_35/cam_high/frame_000070.jpg", "images/robocasa_data/turn_on_stove/episode_35/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/turn_on_stove/episode_35/cam_right_wrist/frame_000070.jpg", "images/robocasa_data/turn_on_stove/episode_35/cam_high/frame_000160.jpg", "images/robocasa_data/turn_on_stove/episode_35/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/turn_on_stove/episode_35/cam_right_wrist/frame_000160.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 39.1, "hop_absolute_value": 39.1, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_19-train-bf_000120-af_000170-plus-0000692", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_32/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_32/cam_high/frame_000343.jpg", "images/robocasa_data/coffee_serve_mug/episode_19/cam_high/frame_000120.jpg", "images/robocasa_data/coffee_serve_mug/episode_19/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/coffee_serve_mug/episode_19/cam_right_wrist/frame_000120.jpg", "images/robocasa_data/coffee_serve_mug/episode_19/cam_high/frame_000170.jpg", "images/robocasa_data/coffee_serve_mug/episode_19/cam_left_wrist/frame_000170.jpg", "images/robocasa_data/coffee_serve_mug/episode_19/cam_right_wrist/frame_000170.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 27.8, "hop_absolute_value": 27.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_microwave_episode_4-train-bf_000030-af_000230-plus-0000827", "task": "turn off microwave", "image": [ "images/robocasa_data/turn_off_microwave/episode_26/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_microwave/episode_26/cam_high/frame_000324.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_high/frame_000030.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_right_wrist/frame_000030.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_high/frame_000230.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_left_wrist/frame_000230.jpg", "images/robocasa_data/turn_off_microwave/episode_4/cam_right_wrist/frame_000230.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 76.9, "hop_absolute_value": 76.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_single_door_episode_28-train-bf_000110-af_000290-plus-0000490", "task": "close single door", "image": [ "images/robocasa_data/close_single_door/episode_17/cam_high/frame_000000.jpg", "images/robocasa_data/close_single_door/episode_17/cam_high/frame_000227.jpg", "images/robocasa_data/close_single_door/episode_28/cam_high/frame_000110.jpg", "images/robocasa_data/close_single_door/episode_28/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/close_single_door/episode_28/cam_right_wrist/frame_000110.jpg", "images/robocasa_data/close_single_door/episode_28/cam_high/frame_000290.jpg", "images/robocasa_data/close_single_door/episode_28/cam_left_wrist/frame_000290.jpg", "images/robocasa_data/close_single_door/episode_28/cam_right_wrist/frame_000290.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_microwave_episode_22-train-bf_000130-af_000220-plus-0000384", "task": "turn on microwave", "image": [ "images/robocasa_data/turn_on_microwave/episode_22/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_microwave/episode_22/cam_high/frame_000256.jpg", "images/robocasa_data/turn_on_microwave/episode_22/cam_high/frame_000130.jpg", "images/robocasa_data/turn_on_microwave/episode_22/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/turn_on_microwave/episode_22/cam_right_wrist/frame_000130.jpg", "images/robocasa_data/turn_on_microwave/episode_22/cam_high/frame_000220.jpg", "images/robocasa_data/turn_on_microwave/episode_22/cam_left_wrist/frame_000220.jpg", "images/robocasa_data/turn_on_microwave/episode_22/cam_right_wrist/frame_000220.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_sink_to_counter_episode_28-train-bf_000040-af_000130-plus-0000290", "task": "pick and place sink to counter", "image": [ "images/robocasa_data/pick_and_place_sink_to_counter/episode_8/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_8/cam_high/frame_000245.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_28/cam_high/frame_000040.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_28/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_28/cam_right_wrist/frame_000040.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_28/cam_high/frame_000130.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_28/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_28/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place sink to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place sink to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 52.9, "hop_absolute_value": 52.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_cab_episode_33-train-bf_000000-af_000210-plus-0000413", "task": "pick and place counter to cab", "image": [ "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_high/frame_000224.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_left_wrist/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_right_wrist/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_high/frame_000210.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_33/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to cab`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to cab`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 95.5, "hop_absolute_value": 95.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_sink_episode_26-train-bf_000120-af_000390-plus-0000455", "task": "pick and place counter to sink", "image": [ "images/robocasa_data/pick_and_place_counter_to_sink/episode_42/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_42/cam_high/frame_000461.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_26/cam_high/frame_000120.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_26/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_26/cam_right_wrist/frame_000120.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_26/cam_high/frame_000390.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_26/cam_left_wrist/frame_000390.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_26/cam_right_wrist/frame_000390.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 67.5, "hop_absolute_value": 67.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_28-train-bf_000200-af_000220-plus-0000669", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_28/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_28/cam_high/frame_000776.jpg", "images/robocasa_data/open_double_door/episode_28/cam_high/frame_000200.jpg", "images/robocasa_data/open_double_door/episode_28/cam_left_wrist/frame_000200.jpg", "images/robocasa_data/open_double_door/episode_28/cam_right_wrist/frame_000200.jpg", "images/robocasa_data/open_double_door/episode_28/cam_high/frame_000220.jpg", "images/robocasa_data/open_double_door/episode_28/cam_left_wrist/frame_000220.jpg", "images/robocasa_data/open_double_door/episode_28/cam_right_wrist/frame_000220.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.5, "hop_absolute_value": 3.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_sink_episode_48-train-bf_000040-af_000420-plus-0000549", "task": "pick and place counter to sink", "image": [ "images/robocasa_data/pick_and_place_counter_to_sink/episode_32/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_32/cam_high/frame_000399.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_48/cam_high/frame_000040.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_48/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_48/cam_right_wrist/frame_000040.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_48/cam_high/frame_000420.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_48/cam_left_wrist/frame_000420.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_48/cam_right_wrist/frame_000420.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 88.4, "hop_absolute_value": 88.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_drawer_episode_403-train-bf_000000-af_000060-plus-0000215", "task": "close drawer", "image": [ "images/robocasa_data/close_drawer/episode_393/cam_high/frame_000000.jpg", "images/robocasa_data/close_drawer/episode_393/cam_high/frame_000215.jpg", "images/robocasa_data/close_drawer/episode_403/cam_high/frame_000000.jpg", "images/robocasa_data/close_drawer/episode_403/cam_left_wrist/frame_000000.jpg", "images/robocasa_data/close_drawer/episode_403/cam_right_wrist/frame_000000.jpg", "images/robocasa_data/close_drawer/episode_403/cam_high/frame_000060.jpg", "images/robocasa_data/close_drawer/episode_403/cam_left_wrist/frame_000060.jpg", "images/robocasa_data/close_drawer/episode_403/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 31.6, "hop_absolute_value": 31.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_microwave_episode_6-train-bf_000060-af_000140-plus-0000876", "task": "pick and place counter to microwave", "image": [ "images/robocasa_data/pick_and_place_counter_to_microwave/episode_39/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_39/cam_high/frame_000474.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_6/cam_high/frame_000060.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_6/cam_left_wrist/frame_000060.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_6/cam_right_wrist/frame_000060.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_6/cam_high/frame_000140.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_6/cam_left_wrist/frame_000140.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_6/cam_right_wrist/frame_000140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 23.5, "hop_absolute_value": 23.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_microwave_episode_42-train-bf_000150-af_000180-plus-0000724", "task": "pick and place counter to microwave", "image": [ "images/robocasa_data/pick_and_place_counter_to_microwave/episode_40/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_40/cam_high/frame_000399.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_42/cam_high/frame_000150.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_42/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_42/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_42/cam_high/frame_000180.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_42/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_42/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_15-train-bf_000030-af_000240-plus-0000471", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_6/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_6/cam_high/frame_000302.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_15/cam_high/frame_000030.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_15/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_15/cam_right_wrist/frame_000030.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_15/cam_high/frame_000240.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_15/cam_left_wrist/frame_000240.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_15/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 95.5, "hop_absolute_value": 95.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_23-train-bf_000150-af_000220-plus-0000337", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_22/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_22/cam_high/frame_000319.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_high/frame_000150.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_high/frame_000220.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_left_wrist/frame_000220.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_right_wrist/frame_000220.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 38.9, "hop_absolute_value": 38.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_4-train-bf_000160-af_000220-plus-0000940", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_26/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_26/cam_high/frame_000673.jpg", "images/robocasa_data/open_double_door/episode_4/cam_high/frame_000160.jpg", "images/robocasa_data/open_double_door/episode_4/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/open_double_door/episode_4/cam_right_wrist/frame_000160.jpg", "images/robocasa_data/open_double_door/episode_4/cam_high/frame_000220.jpg", "images/robocasa_data/open_double_door/episode_4/cam_left_wrist/frame_000220.jpg", "images/robocasa_data/open_double_door/episode_4/cam_right_wrist/frame_000220.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.9, "hop_absolute_value": 10.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_11-train-bf_000100-af_000280-plus-0000486", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_1/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_1/cam_high/frame_000472.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_11/cam_high/frame_000100.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_11/cam_left_wrist/frame_000100.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_11/cam_right_wrist/frame_000100.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_11/cam_high/frame_000280.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_11/cam_left_wrist/frame_000280.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_11/cam_right_wrist/frame_000280.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 90.0, "hop_absolute_value": 90.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_sink_episode_41-train-bf_000290-af_000300-plus-0000763", "task": "pick and place counter to sink", "image": [ "images/robocasa_data/pick_and_place_counter_to_sink/episode_40/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_40/cam_high/frame_000406.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_41/cam_high/frame_000290.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_41/cam_left_wrist/frame_000290.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_41/cam_right_wrist/frame_000290.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_41/cam_high/frame_000300.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_41/cam_left_wrist/frame_000300.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_41/cam_right_wrist/frame_000300.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.3, "hop_absolute_value": 4.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_sink_to_counter_episode_35-train-bf_000070-af_000120-plus-0000262", "task": "pick and place sink to counter", "image": [ "images/robocasa_data/pick_and_place_sink_to_counter/episode_9/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_9/cam_high/frame_000296.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_35/cam_high/frame_000070.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_35/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_35/cam_right_wrist/frame_000070.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_35/cam_high/frame_000120.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_35/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_35/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place sink to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place sink to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 29.4, "hop_absolute_value": 29.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_stove_episode_10-train-bf_000200-af_000240-plus-0000325", "task": "pick and place counter to stove", "image": [ "images/robocasa_data/pick_and_place_counter_to_stove/episode_2/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_2/cam_high/frame_000347.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_10/cam_high/frame_000200.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_10/cam_left_wrist/frame_000200.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_10/cam_right_wrist/frame_000200.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_10/cam_high/frame_000240.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_10/cam_left_wrist/frame_000240.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_10/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 28.6, "hop_absolute_value": 28.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_30-train-bf_000110-af_000680-plus-0000598", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_32/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_32/cam_high/frame_000749.jpg", "images/robocasa_data/open_double_door/episode_30/cam_high/frame_000110.jpg", "images/robocasa_data/open_double_door/episode_30/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/open_double_door/episode_30/cam_right_wrist/frame_000110.jpg", "images/robocasa_data/open_double_door/episode_30/cam_high/frame_000680.jpg", "images/robocasa_data/open_double_door/episode_30/cam_left_wrist/frame_000680.jpg", "images/robocasa_data/open_double_door/episode_30/cam_right_wrist/frame_000680.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 95.0, "hop_absolute_value": 95.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_stove_episode_23-train-bf_000280-af_000316-plus-0000512", "task": "pick and place counter to stove", "image": [ "images/robocasa_data/pick_and_place_counter_to_stove/episode_16/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_16/cam_high/frame_000292.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_23/cam_high/frame_000280.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_23/cam_left_wrist/frame_000280.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_23/cam_right_wrist/frame_000280.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_23/cam_high/frame_000316.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_23/cam_left_wrist/frame_000316.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_23/cam_right_wrist/frame_000316.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_microwave_episode_2-train-bf_000130-af_000240-plus-0000421", "task": "turn on microwave", "image": [ "images/robocasa_data/turn_on_microwave/episode_0/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_microwave/episode_0/cam_high/frame_000273.jpg", "images/robocasa_data/turn_on_microwave/episode_2/cam_high/frame_000130.jpg", "images/robocasa_data/turn_on_microwave/episode_2/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/turn_on_microwave/episode_2/cam_right_wrist/frame_000130.jpg", "images/robocasa_data/turn_on_microwave/episode_2/cam_high/frame_000240.jpg", "images/robocasa_data/turn_on_microwave/episode_2/cam_left_wrist/frame_000240.jpg", "images/robocasa_data/turn_on_microwave/episode_2/cam_right_wrist/frame_000240.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 78.6, "hop_absolute_value": 78.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_12-train-bf_000010-af_000140-plus-0000351", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_23/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_high/frame_000336.jpg", "images/robocasa_data/coffee_serve_mug/episode_12/cam_high/frame_000010.jpg", "images/robocasa_data/coffee_serve_mug/episode_12/cam_left_wrist/frame_000010.jpg", "images/robocasa_data/coffee_serve_mug/episode_12/cam_right_wrist/frame_000010.jpg", "images/robocasa_data/coffee_serve_mug/episode_12/cam_high/frame_000140.jpg", "images/robocasa_data/coffee_serve_mug/episode_12/cam_left_wrist/frame_000140.jpg", "images/robocasa_data/coffee_serve_mug/episode_12/cam_right_wrist/frame_000140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 30.2, "hop_absolute_value": 30.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_microwave_episode_31-train-bf_000070-af_000120-plus-0000266", "task": "turn off microwave", "image": [ "images/robocasa_data/turn_off_microwave/episode_31/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_microwave/episode_31/cam_high/frame_000253.jpg", "images/robocasa_data/turn_off_microwave/episode_31/cam_high/frame_000070.jpg", "images/robocasa_data/turn_off_microwave/episode_31/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/turn_off_microwave/episode_31/cam_right_wrist/frame_000070.jpg", "images/robocasa_data/turn_off_microwave/episode_31/cam_high/frame_000120.jpg", "images/robocasa_data/turn_off_microwave/episode_31/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/turn_off_microwave/episode_31/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 27.8, "hop_absolute_value": 27.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_setup_mug_episode_42-train-bf_000170-af_000230-plus-0000908", "task": "coffee setup mug", "image": [ "images/robocasa_data/coffee_setup_mug/episode_42/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_42/cam_high/frame_000370.jpg", "images/robocasa_data/coffee_setup_mug/episode_42/cam_high/frame_000170.jpg", "images/robocasa_data/coffee_setup_mug/episode_42/cam_left_wrist/frame_000170.jpg", "images/robocasa_data/coffee_setup_mug/episode_42/cam_right_wrist/frame_000170.jpg", "images/robocasa_data/coffee_setup_mug/episode_42/cam_high/frame_000230.jpg", "images/robocasa_data/coffee_setup_mug/episode_42/cam_left_wrist/frame_000230.jpg", "images/robocasa_data/coffee_setup_mug/episode_42/cam_right_wrist/frame_000230.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee setup mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee setup mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 30.0, "hop_absolute_value": 30.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_sink_faucet_episode_15-train-bf_000030-af_000040-plus-0000279", "task": "turn on sink faucet", "image": [ "images/robocasa_data/turn_on_sink_faucet/episode_15/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_15/cam_high/frame_000333.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_15/cam_high/frame_000030.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_15/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_15/cam_right_wrist/frame_000030.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_15/cam_high/frame_000040.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_15/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_15/cam_right_wrist/frame_000040.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.3, "hop_absolute_value": 3.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_microwave_to_counter_episode_35-train-bf_000020-af_000140-plus-0000656", "task": "pick and place microwave to counter", "image": [ "images/robocasa_data/pick_and_place_microwave_to_counter/episode_10/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_10/cam_high/frame_000333.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_35/cam_high/frame_000020.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_35/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_35/cam_right_wrist/frame_000020.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_35/cam_high/frame_000140.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_35/cam_left_wrist/frame_000140.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_35/cam_right_wrist/frame_000140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place microwave to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place microwave to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_single_door_episode_27-train-bf_000090-af_000210-plus-0000393", "task": "close single door", "image": [ "images/robocasa_data/close_single_door/episode_7/cam_high/frame_000000.jpg", "images/robocasa_data/close_single_door/episode_7/cam_high/frame_000311.jpg", "images/robocasa_data/close_single_door/episode_27/cam_high/frame_000090.jpg", "images/robocasa_data/close_single_door/episode_27/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/close_single_door/episode_27/cam_right_wrist/frame_000090.jpg", "images/robocasa_data/close_single_door/episode_27/cam_high/frame_000210.jpg", "images/robocasa_data/close_single_door/episode_27/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/close_single_door/episode_27/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.6, "hop_absolute_value": 70.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_sink_episode_3-train-bf_000370-af_000380-plus-0000308", "task": "pick and place counter to sink", "image": [ "images/robocasa_data/pick_and_place_counter_to_sink/episode_3/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_3/cam_high/frame_000498.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_3/cam_high/frame_000370.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_3/cam_left_wrist/frame_000370.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_3/cam_right_wrist/frame_000370.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_3/cam_high/frame_000380.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_3/cam_left_wrist/frame_000380.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_3/cam_right_wrist/frame_000380.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 8.3, "hop_absolute_value": 8.3, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_sink_spout_episode_3-train-bf_000120-af_000130-plus-0000159", "task": "turn sink spout", "image": [ "images/robocasa_data/turn_sink_spout/episode_3/cam_high/frame_000000.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_high/frame_000160.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_high/frame_000120.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_right_wrist/frame_000120.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_high/frame_000130.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/turn_sink_spout/episode_3/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn sink spout`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn sink spout`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_stove_episode_43-train-bf_000080-af_000090-plus-0000583", "task": "turn on stove", "image": [ "images/robocasa_data/turn_on_stove/episode_25/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_stove/episode_25/cam_high/frame_000252.jpg", "images/robocasa_data/turn_on_stove/episode_43/cam_high/frame_000080.jpg", "images/robocasa_data/turn_on_stove/episode_43/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/turn_on_stove/episode_43/cam_right_wrist/frame_000080.jpg", "images/robocasa_data/turn_on_stove/episode_43/cam_high/frame_000090.jpg", "images/robocasa_data/turn_on_stove/episode_43/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/turn_on_stove/episode_43/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.5, "hop_absolute_value": 4.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_stove_episode_35-train-bf_000090-af_000100-plus-0000123", "task": "turn off stove", "image": [ "images/robocasa_data/turn_off_stove/episode_12/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_stove/episode_12/cam_high/frame_000143.jpg", "images/robocasa_data/turn_off_stove/episode_35/cam_high/frame_000090.jpg", "images/robocasa_data/turn_off_stove/episode_35/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/turn_off_stove/episode_35/cam_right_wrist/frame_000090.jpg", "images/robocasa_data/turn_off_stove/episode_35/cam_high/frame_000100.jpg", "images/robocasa_data/turn_off_stove/episode_35/cam_left_wrist/frame_000100.jpg", "images/robocasa_data/turn_off_stove/episode_35/cam_right_wrist/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_microwave_episode_13-train-bf_000170-af_000030-minus-0000049", "task": "turn on microwave", "image": [ "images/robocasa_data/turn_on_microwave/episode_13/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_microwave/episode_13/cam_high/frame_000193.jpg", "images/robocasa_data/turn_on_microwave/episode_13/cam_high/frame_000170.jpg", "images/robocasa_data/turn_on_microwave/episode_13/cam_left_wrist/frame_000170.jpg", "images/robocasa_data/turn_on_microwave/episode_13/cam_right_wrist/frame_000170.jpg", "images/robocasa_data/turn_on_microwave/episode_13/cam_high/frame_000030.jpg", "images/robocasa_data/turn_on_microwave/episode_13/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/turn_on_microwave/episode_13/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -82.4, "hop_absolute_value": 82.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_setup_mug_episode_56-train-bf_000080-af_000060-minus-0000190", "task": "coffee setup mug", "image": [ "images/robocasa_data/coffee_setup_mug/episode_61/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_61/cam_high/frame_000275.jpg", "images/robocasa_data/coffee_setup_mug/episode_56/cam_high/frame_000080.jpg", "images/robocasa_data/coffee_setup_mug/episode_56/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/coffee_setup_mug/episode_56/cam_right_wrist/frame_000080.jpg", "images/robocasa_data/coffee_setup_mug/episode_56/cam_high/frame_000060.jpg", "images/robocasa_data/coffee_setup_mug/episode_56/cam_left_wrist/frame_000060.jpg", "images/robocasa_data/coffee_setup_mug/episode_56/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee setup mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee setup mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_microwave_to_counter_episode_19-train-bf_000210-af_000150-minus-0000736", "task": "pick and place microwave to counter", "image": [ "images/robocasa_data/pick_and_place_microwave_to_counter/episode_3/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_3/cam_high/frame_000332.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_19/cam_high/frame_000210.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_19/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_19/cam_right_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_19/cam_high/frame_000150.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_19/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_19/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place microwave to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place microwave to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -28.6, "hop_absolute_value": 28.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_microwave_episode_2-train-bf_000273-af_000110-minus-0000155", "task": "turn on microwave", "image": [ "images/robocasa_data/turn_on_microwave/episode_14/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_microwave/episode_14/cam_high/frame_000216.jpg", "images/robocasa_data/turn_on_microwave/episode_2/cam_high/frame_000273.jpg", "images/robocasa_data/turn_on_microwave/episode_2/cam_left_wrist/frame_000273.jpg", "images/robocasa_data/turn_on_microwave/episode_2/cam_right_wrist/frame_000273.jpg", "images/robocasa_data/turn_on_microwave/episode_2/cam_high/frame_000110.jpg", "images/robocasa_data/turn_on_microwave/episode_2/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/turn_on_microwave/episode_2/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -59.3, "hop_absolute_value": 59.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_microwave_episode_27-train-bf_000020-af_000010-minus-0000148", "task": "turn on microwave", "image": [ "images/robocasa_data/turn_on_microwave/episode_29/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_microwave/episode_29/cam_high/frame_000184.jpg", "images/robocasa_data/turn_on_microwave/episode_27/cam_high/frame_000020.jpg", "images/robocasa_data/turn_on_microwave/episode_27/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/turn_on_microwave/episode_27/cam_right_wrist/frame_000020.jpg", "images/robocasa_data/turn_on_microwave/episode_27/cam_high/frame_000010.jpg", "images/robocasa_data/turn_on_microwave/episode_27/cam_left_wrist/frame_000010.jpg", "images/robocasa_data/turn_on_microwave/episode_27/cam_right_wrist/frame_000010.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_microwave_episode_20-train-bf_000150-af_000010-minus-0000643", "task": "turn off microwave", "image": [ "images/robocasa_data/turn_off_microwave/episode_13/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_microwave/episode_13/cam_high/frame_000209.jpg", "images/robocasa_data/turn_off_microwave/episode_20/cam_high/frame_000150.jpg", "images/robocasa_data/turn_off_microwave/episode_20/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/turn_off_microwave/episode_20/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/turn_off_microwave/episode_20/cam_high/frame_000010.jpg", "images/robocasa_data/turn_off_microwave/episode_20/cam_left_wrist/frame_000010.jpg", "images/robocasa_data/turn_off_microwave/episode_20/cam_right_wrist/frame_000010.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -93.3, "hop_absolute_value": 93.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_microwave_episode_14-train-bf_000260-af_000250-minus-0000296", "task": "pick and place counter to microwave", "image": [ "images/robocasa_data/pick_and_place_counter_to_microwave/episode_14/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_14/cam_high/frame_000429.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_14/cam_high/frame_000260.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_14/cam_left_wrist/frame_000260.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_14/cam_right_wrist/frame_000260.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_14/cam_high/frame_000250.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_14/cam_left_wrist/frame_000250.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_14/cam_right_wrist/frame_000250.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.8, "hop_absolute_value": 3.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_stove_episode_15-train-bf_000260-af_000040-minus-0000052", "task": "pick and place counter to stove", "image": [ "images/robocasa_data/pick_and_place_counter_to_stove/episode_15/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_15/cam_high/frame_000353.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_15/cam_high/frame_000260.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_15/cam_left_wrist/frame_000260.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_15/cam_right_wrist/frame_000260.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_15/cam_high/frame_000040.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_15/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_15/cam_right_wrist/frame_000040.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -84.6, "hop_absolute_value": 84.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_stove_episode_29-train-bf_000250-af_000090-minus-0000136", "task": "pick and place counter to stove", "image": [ "images/robocasa_data/pick_and_place_counter_to_stove/episode_15/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_15/cam_high/frame_000353.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_29/cam_high/frame_000250.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_29/cam_left_wrist/frame_000250.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_29/cam_right_wrist/frame_000250.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_29/cam_high/frame_000090.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_29/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_29/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -64.0, "hop_absolute_value": 64.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_36-train-bf_000070-af_000040-minus-0000201", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_30/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_30/cam_high/frame_000373.jpg", "images/robocasa_data/coffee_serve_mug/episode_36/cam_high/frame_000070.jpg", "images/robocasa_data/coffee_serve_mug/episode_36/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/coffee_serve_mug/episode_36/cam_right_wrist/frame_000070.jpg", "images/robocasa_data/coffee_serve_mug/episode_36/cam_high/frame_000040.jpg", "images/robocasa_data/coffee_serve_mug/episode_36/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/coffee_serve_mug/episode_36/cam_right_wrist/frame_000040.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -42.9, "hop_absolute_value": 42.9, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_microwave_episode_46-train-bf_000420-af_000050-minus-0000071", "task": "pick and place counter to microwave", "image": [ "images/robocasa_data/pick_and_place_counter_to_microwave/episode_40/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_40/cam_high/frame_000399.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_46/cam_high/frame_000420.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_46/cam_left_wrist/frame_000420.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_46/cam_right_wrist/frame_000420.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_46/cam_high/frame_000050.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_46/cam_left_wrist/frame_000050.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_46/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -88.1, "hop_absolute_value": 88.1, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_double_door_episode_29-train-bf_000320-af_000120-minus-0000146", "task": "close double door", "image": [ "images/robocasa_data/close_double_door/episode_24/cam_high/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_24/cam_high/frame_000375.jpg", "images/robocasa_data/close_double_door/episode_29/cam_high/frame_000320.jpg", "images/robocasa_data/close_double_door/episode_29/cam_left_wrist/frame_000320.jpg", "images/robocasa_data/close_double_door/episode_29/cam_right_wrist/frame_000320.jpg", "images/robocasa_data/close_double_door/episode_29/cam_high/frame_000120.jpg", "images/robocasa_data/close_double_door/episode_29/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/close_double_door/episode_29/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_single_door_episode_19-train-bf_000210-af_000010-minus-0000641", "task": "close single door", "image": [ "images/robocasa_data/close_single_door/episode_16/cam_high/frame_000000.jpg", "images/robocasa_data/close_single_door/episode_16/cam_high/frame_000263.jpg", "images/robocasa_data/close_single_door/episode_19/cam_high/frame_000210.jpg", "images/robocasa_data/close_single_door/episode_19/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/close_single_door/episode_19/cam_right_wrist/frame_000210.jpg", "images/robocasa_data/close_single_door/episode_19/cam_high/frame_000010.jpg", "images/robocasa_data/close_single_door/episode_19/cam_left_wrist/frame_000010.jpg", "images/robocasa_data/close_single_door/episode_19/cam_right_wrist/frame_000010.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -95.2, "hop_absolute_value": 95.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_double_door_episode_32-train-bf_000050-af_000040-minus-0000759", "task": "close double door", "image": [ "images/robocasa_data/close_double_door/episode_19/cam_high/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_19/cam_high/frame_000399.jpg", "images/robocasa_data/close_double_door/episode_32/cam_high/frame_000050.jpg", "images/robocasa_data/close_double_door/episode_32/cam_left_wrist/frame_000050.jpg", "images/robocasa_data/close_double_door/episode_32/cam_right_wrist/frame_000050.jpg", "images/robocasa_data/close_double_door/episode_32/cam_high/frame_000040.jpg", "images/robocasa_data/close_double_door/episode_32/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/close_double_door/episode_32/cam_right_wrist/frame_000040.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_cab_episode_9-train-bf_000270-af_000160-minus-0000201", "task": "pick and place counter to cab", "image": [ "images/robocasa_data/pick_and_place_counter_to_cab/episode_9/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_9/cam_high/frame_000286.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_9/cam_high/frame_000270.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_9/cam_left_wrist/frame_000270.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_9/cam_right_wrist/frame_000270.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_9/cam_high/frame_000160.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_9/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/pick_and_place_counter_to_cab/episode_9/cam_right_wrist/frame_000160.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to cab`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to cab`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.7, "hop_absolute_value": 40.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_microwave_episode_22-train-bf_000280-af_000080-minus-0000127", "task": "pick and place counter to microwave", "image": [ "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_high/frame_000461.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_high/frame_000280.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_left_wrist/frame_000280.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_right_wrist/frame_000280.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_high/frame_000080.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_22/cam_right_wrist/frame_000080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_stove_to_counter_episode_18-train-bf_000210-af_000100-minus-0000156", "task": "pick and place stove to counter", "image": [ "images/robocasa_data/pick_and_place_stove_to_counter/episode_33/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_33/cam_high/frame_000262.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_18/cam_high/frame_000210.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_18/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_18/cam_right_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_18/cam_high/frame_000100.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_18/cam_left_wrist/frame_000100.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_18/cam_right_wrist/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place stove to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place stove to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -52.4, "hop_absolute_value": 52.4, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_sink_faucet_episode_19-train-bf_000120-af_000050-minus-0000499", "task": "turn on sink faucet", "image": [ "images/robocasa_data/turn_on_sink_faucet/episode_15/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_15/cam_high/frame_000333.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_high/frame_000120.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_right_wrist/frame_000120.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_high/frame_000050.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_left_wrist/frame_000050.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -58.3, "hop_absolute_value": 58.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_single_door_episode_0-train-bf_000220-af_000080-minus-0000117", "task": "open single door", "image": [ "images/robocasa_data/open_single_door/episode_28/cam_high/frame_000000.jpg", "images/robocasa_data/open_single_door/episode_28/cam_high/frame_000241.jpg", "images/robocasa_data/open_single_door/episode_0/cam_high/frame_000220.jpg", "images/robocasa_data/open_single_door/episode_0/cam_left_wrist/frame_000220.jpg", "images/robocasa_data/open_single_door/episode_0/cam_right_wrist/frame_000220.jpg", "images/robocasa_data/open_single_door/episode_0/cam_high/frame_000080.jpg", "images/robocasa_data/open_single_door/episode_0/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/open_single_door/episode_0/cam_right_wrist/frame_000080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -63.6, "hop_absolute_value": 63.6, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_21-train-bf_000280-af_000110-minus-0000789", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_3/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_3/cam_high/frame_000296.jpg", "images/robocasa_data/coffee_serve_mug/episode_21/cam_high/frame_000280.jpg", "images/robocasa_data/coffee_serve_mug/episode_21/cam_left_wrist/frame_000280.jpg", "images/robocasa_data/coffee_serve_mug/episode_21/cam_right_wrist/frame_000280.jpg", "images/robocasa_data/coffee_serve_mug/episode_21/cam_high/frame_000110.jpg", "images/robocasa_data/coffee_serve_mug/episode_21/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/coffee_serve_mug/episode_21/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -60.7, "hop_absolute_value": 60.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_sink_to_counter_episode_24-train-bf_000210-af_000040-minus-0000076", "task": "pick and place sink to counter", "image": [ "images/robocasa_data/pick_and_place_sink_to_counter/episode_9/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_9/cam_high/frame_000296.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_24/cam_high/frame_000210.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_24/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_24/cam_right_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_24/cam_high/frame_000040.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_24/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_24/cam_right_wrist/frame_000040.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place sink to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place sink to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -81.0, "hop_absolute_value": 81.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_stove_episode_6-train-bf_000160-af_000030-minus-0000070", "task": "pick and place counter to stove", "image": [ "images/robocasa_data/pick_and_place_counter_to_stove/episode_34/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_34/cam_high/frame_000255.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_6/cam_high/frame_000160.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_6/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_6/cam_right_wrist/frame_000160.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_6/cam_high/frame_000030.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_6/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_6/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -81.2, "hop_absolute_value": 81.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_sink_faucet_episode_19-train-bf_000190-af_000160-minus-0000444", "task": "turn on sink faucet", "image": [ "images/robocasa_data/turn_on_sink_faucet/episode_9/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_9/cam_high/frame_000334.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_high/frame_000190.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_left_wrist/frame_000190.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_right_wrist/frame_000190.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_high/frame_000160.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_19/cam_right_wrist/frame_000160.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -15.8, "hop_absolute_value": 15.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_microwave_to_counter_episode_21-train-bf_000230-af_000110-minus-0000146", "task": "pick and place microwave to counter", "image": [ "images/robocasa_data/pick_and_place_microwave_to_counter/episode_37/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_37/cam_high/frame_000368.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_21/cam_high/frame_000230.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_21/cam_left_wrist/frame_000230.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_21/cam_right_wrist/frame_000230.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_21/cam_high/frame_000110.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_21/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_21/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place microwave to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place microwave to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -52.2, "hop_absolute_value": 52.2, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_5-train-bf_000310-af_000070-minus-0000094", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_30/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_30/cam_high/frame_000373.jpg", "images/robocasa_data/coffee_serve_mug/episode_5/cam_high/frame_000310.jpg", "images/robocasa_data/coffee_serve_mug/episode_5/cam_left_wrist/frame_000310.jpg", "images/robocasa_data/coffee_serve_mug/episode_5/cam_right_wrist/frame_000310.jpg", "images/robocasa_data/coffee_serve_mug/episode_5/cam_high/frame_000070.jpg", "images/robocasa_data/coffee_serve_mug/episode_5/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/coffee_serve_mug/episode_5/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -77.4, "hop_absolute_value": 77.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_drawer_episode_385-train-bf_000140-af_000110-minus-0000332", "task": "close drawer", "image": [ "images/robocasa_data/close_drawer/episode_385/cam_high/frame_000000.jpg", "images/robocasa_data/close_drawer/episode_385/cam_high/frame_000187.jpg", "images/robocasa_data/close_drawer/episode_385/cam_high/frame_000140.jpg", "images/robocasa_data/close_drawer/episode_385/cam_left_wrist/frame_000140.jpg", "images/robocasa_data/close_drawer/episode_385/cam_right_wrist/frame_000140.jpg", "images/robocasa_data/close_drawer/episode_385/cam_high/frame_000110.jpg", "images/robocasa_data/close_drawer/episode_385/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/close_drawer/episode_385/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -21.4, "hop_absolute_value": 21.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_11-train-bf_000210-af_000190-minus-0000814", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_30/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_30/cam_high/frame_000712.jpg", "images/robocasa_data/open_double_door/episode_11/cam_high/frame_000210.jpg", "images/robocasa_data/open_double_door/episode_11/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/open_double_door/episode_11/cam_right_wrist/frame_000210.jpg", "images/robocasa_data/open_double_door/episode_11/cam_high/frame_000190.jpg", "images/robocasa_data/open_double_door/episode_11/cam_left_wrist/frame_000190.jpg", "images/robocasa_data/open_double_door/episode_11/cam_right_wrist/frame_000190.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -9.5, "hop_absolute_value": 9.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_sink_faucet_episode_35-train-bf_000030-af_000000-minus-0000002", "task": "turn on sink faucet", "image": [ "images/robocasa_data/turn_on_sink_faucet/episode_35/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_35/cam_high/frame_000245.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_35/cam_high/frame_000030.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_35/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_35/cam_right_wrist/frame_000030.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_35/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_35/cam_left_wrist/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_35/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_36-train-bf_000150-af_000090-minus-0000198", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_21/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_21/cam_high/frame_000297.jpg", "images/robocasa_data/coffee_serve_mug/episode_36/cam_high/frame_000150.jpg", "images/robocasa_data/coffee_serve_mug/episode_36/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/coffee_serve_mug/episode_36/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/coffee_serve_mug/episode_36/cam_high/frame_000090.jpg", "images/robocasa_data/coffee_serve_mug/episode_36/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/coffee_serve_mug/episode_36/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_single_door_episode_29-train-bf_000150-af_000090-minus-0000180", "task": "open single door", "image": [ "images/robocasa_data/open_single_door/episode_18/cam_high/frame_000000.jpg", "images/robocasa_data/open_single_door/episode_18/cam_high/frame_000247.jpg", "images/robocasa_data/open_single_door/episode_29/cam_high/frame_000150.jpg", "images/robocasa_data/open_single_door/episode_29/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/open_single_door/episode_29/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/open_single_door/episode_29/cam_high/frame_000090.jpg", "images/robocasa_data/open_single_door/episode_29/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/open_single_door/episode_29/cam_right_wrist/frame_000090.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_35-train-bf_000670-af_000290-minus-0000200", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_6/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_6/cam_high/frame_000727.jpg", "images/robocasa_data/open_double_door/episode_35/cam_high/frame_000670.jpg", "images/robocasa_data/open_double_door/episode_35/cam_left_wrist/frame_000670.jpg", "images/robocasa_data/open_double_door/episode_35/cam_right_wrist/frame_000670.jpg", "images/robocasa_data/open_double_door/episode_35/cam_high/frame_000290.jpg", "images/robocasa_data/open_double_door/episode_35/cam_left_wrist/frame_000290.jpg", "images/robocasa_data/open_double_door/episode_35/cam_right_wrist/frame_000290.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -56.7, "hop_absolute_value": 56.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_30-train-bf_000150-af_000070-minus-0000172", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_17/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_17/cam_high/frame_000779.jpg", "images/robocasa_data/open_double_door/episode_30/cam_high/frame_000150.jpg", "images/robocasa_data/open_double_door/episode_30/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/open_double_door/episode_30/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/open_double_door/episode_30/cam_high/frame_000070.jpg", "images/robocasa_data/open_double_door/episode_30/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/open_double_door/episode_30/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -53.3, "hop_absolute_value": 53.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_single_door_episode_15-train-bf_000130-af_000020-minus-0000037", "task": "open single door", "image": [ "images/robocasa_data/open_single_door/episode_14/cam_high/frame_000000.jpg", "images/robocasa_data/open_single_door/episode_14/cam_high/frame_000337.jpg", "images/robocasa_data/open_single_door/episode_15/cam_high/frame_000130.jpg", "images/robocasa_data/open_single_door/episode_15/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/open_single_door/episode_15/cam_right_wrist/frame_000130.jpg", "images/robocasa_data/open_single_door/episode_15/cam_high/frame_000020.jpg", "images/robocasa_data/open_single_door/episode_15/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/open_single_door/episode_15/cam_right_wrist/frame_000020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -84.6, "hop_absolute_value": 84.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_stove_episode_7-train-bf_000400-af_000320-minus-0000739", "task": "pick and place counter to stove", "image": [ "images/robocasa_data/pick_and_place_counter_to_stove/episode_7/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_7/cam_high/frame_000423.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_7/cam_high/frame_000400.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_7/cam_left_wrist/frame_000400.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_7/cam_right_wrist/frame_000400.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_7/cam_high/frame_000320.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_7/cam_left_wrist/frame_000320.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_7/cam_right_wrist/frame_000320.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_microwave_to_counter_episode_15-train-bf_000250-af_000180-minus-0000761", "task": "pick and place microwave to counter", "image": [ "images/robocasa_data/pick_and_place_microwave_to_counter/episode_23/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_23/cam_high/frame_000243.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_high/frame_000250.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_left_wrist/frame_000250.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_right_wrist/frame_000250.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_high/frame_000180.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place microwave to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place microwave to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -28.0, "hop_absolute_value": 28.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_stove_to_counter_episode_21-train-bf_000190-af_000130-minus-0000213", "task": "pick and place stove to counter", "image": [ "images/robocasa_data/pick_and_place_stove_to_counter/episode_8/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_8/cam_high/frame_000208.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_21/cam_high/frame_000190.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_21/cam_left_wrist/frame_000190.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_21/cam_right_wrist/frame_000190.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_21/cam_high/frame_000130.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_21/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_21/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place stove to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place stove to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -31.6, "hop_absolute_value": 31.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_microwave_episode_20-train-bf_000170-af_000120-minus-0000183", "task": "turn on microwave", "image": [ "images/robocasa_data/turn_on_microwave/episode_20/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_microwave/episode_20/cam_high/frame_000243.jpg", "images/robocasa_data/turn_on_microwave/episode_20/cam_high/frame_000170.jpg", "images/robocasa_data/turn_on_microwave/episode_20/cam_left_wrist/frame_000170.jpg", "images/robocasa_data/turn_on_microwave/episode_20/cam_right_wrist/frame_000170.jpg", "images/robocasa_data/turn_on_microwave/episode_20/cam_high/frame_000120.jpg", "images/robocasa_data/turn_on_microwave/episode_20/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/turn_on_microwave/episode_20/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -29.4, "hop_absolute_value": 29.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_26-train-bf_000420-af_000400-minus-0000772", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_5/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_5/cam_high/frame_000765.jpg", "images/robocasa_data/open_double_door/episode_26/cam_high/frame_000420.jpg", "images/robocasa_data/open_double_door/episode_26/cam_left_wrist/frame_000420.jpg", "images/robocasa_data/open_double_door/episode_26/cam_right_wrist/frame_000420.jpg", "images/robocasa_data/open_double_door/episode_26/cam_high/frame_000400.jpg", "images/robocasa_data/open_double_door/episode_26/cam_left_wrist/frame_000400.jpg", "images/robocasa_data/open_double_door/episode_26/cam_right_wrist/frame_000400.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.8, "hop_absolute_value": 4.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_microwave_episode_49-train-bf_000400-af_000180-minus-0000196", "task": "pick and place counter to microwave", "image": [ "images/robocasa_data/pick_and_place_counter_to_microwave/episode_18/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_18/cam_high/frame_000452.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_49/cam_high/frame_000400.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_49/cam_left_wrist/frame_000400.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_49/cam_right_wrist/frame_000400.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_49/cam_high/frame_000180.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_49/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_49/cam_right_wrist/frame_000180.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -55.0, "hop_absolute_value": 55.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_drawer_episode_375-train-bf_000170-af_000120-minus-0000148", "task": "close drawer", "image": [ "images/robocasa_data/close_drawer/episode_380/cam_high/frame_000000.jpg", "images/robocasa_data/close_drawer/episode_380/cam_high/frame_000201.jpg", "images/robocasa_data/close_drawer/episode_375/cam_high/frame_000170.jpg", "images/robocasa_data/close_drawer/episode_375/cam_left_wrist/frame_000170.jpg", "images/robocasa_data/close_drawer/episode_375/cam_right_wrist/frame_000170.jpg", "images/robocasa_data/close_drawer/episode_375/cam_high/frame_000120.jpg", "images/robocasa_data/close_drawer/episode_375/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/close_drawer/episode_375/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -29.4, "hop_absolute_value": 29.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_sink_spout_episode_15-train-bf_000120-af_000100-minus-0000149", "task": "turn sink spout", "image": [ "images/robocasa_data/turn_sink_spout/episode_0/cam_high/frame_000000.jpg", "images/robocasa_data/turn_sink_spout/episode_0/cam_high/frame_000114.jpg", "images/robocasa_data/turn_sink_spout/episode_15/cam_high/frame_000120.jpg", "images/robocasa_data/turn_sink_spout/episode_15/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/turn_sink_spout/episode_15/cam_right_wrist/frame_000120.jpg", "images/robocasa_data/turn_sink_spout/episode_15/cam_high/frame_000100.jpg", "images/robocasa_data/turn_sink_spout/episode_15/cam_left_wrist/frame_000100.jpg", "images/robocasa_data/turn_sink_spout/episode_15/cam_right_wrist/frame_000100.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn sink spout`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn sink spout`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_drawer_episode_2465-train-bf_000040-af_000000-minus-0000003", "task": "open drawer", "image": [ "images/robocasa_data/open_drawer/episode_2475/cam_high/frame_000000.jpg", "images/robocasa_data/open_drawer/episode_2475/cam_high/frame_000254.jpg", "images/robocasa_data/open_drawer/episode_2465/cam_high/frame_000040.jpg", "images/robocasa_data/open_drawer/episode_2465/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/open_drawer/episode_2465/cam_right_wrist/frame_000040.jpg", "images/robocasa_data/open_drawer/episode_2465/cam_high/frame_000000.jpg", "images/robocasa_data/open_drawer/episode_2465/cam_left_wrist/frame_000000.jpg", "images/robocasa_data/open_drawer/episode_2465/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_microwave_episode_1-train-bf_000330-af_000310-minus-0000655", "task": "pick and place counter to microwave", "image": [ "images/robocasa_data/pick_and_place_counter_to_microwave/episode_24/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_24/cam_high/frame_000401.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_1/cam_high/frame_000330.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_1/cam_left_wrist/frame_000330.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_1/cam_right_wrist/frame_000330.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_1/cam_high/frame_000310.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_1/cam_left_wrist/frame_000310.jpg", "images/robocasa_data/pick_and_place_counter_to_microwave/episode_1/cam_right_wrist/frame_000310.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.1, "hop_absolute_value": 6.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_single_door_episode_0-train-bf_000100-af_000060-minus-0000164", "task": "open single door", "image": [ "images/robocasa_data/open_single_door/episode_19/cam_high/frame_000000.jpg", "images/robocasa_data/open_single_door/episode_19/cam_high/frame_000313.jpg", "images/robocasa_data/open_single_door/episode_0/cam_high/frame_000100.jpg", "images/robocasa_data/open_single_door/episode_0/cam_left_wrist/frame_000100.jpg", "images/robocasa_data/open_single_door/episode_0/cam_right_wrist/frame_000100.jpg", "images/robocasa_data/open_single_door/episode_0/cam_high/frame_000060.jpg", "images/robocasa_data/open_single_door/episode_0/cam_left_wrist/frame_000060.jpg", "images/robocasa_data/open_single_door/episode_0/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_single_door_episode_15-train-bf_000060-af_000010-minus-0000051", "task": "open single door", "image": [ "images/robocasa_data/open_single_door/episode_28/cam_high/frame_000000.jpg", "images/robocasa_data/open_single_door/episode_28/cam_high/frame_000241.jpg", "images/robocasa_data/open_single_door/episode_15/cam_high/frame_000060.jpg", "images/robocasa_data/open_single_door/episode_15/cam_left_wrist/frame_000060.jpg", "images/robocasa_data/open_single_door/episode_15/cam_right_wrist/frame_000060.jpg", "images/robocasa_data/open_single_door/episode_15/cam_high/frame_000010.jpg", "images/robocasa_data/open_single_door/episode_15/cam_left_wrist/frame_000010.jpg", "images/robocasa_data/open_single_door/episode_15/cam_right_wrist/frame_000010.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_double_door_episode_35-train-bf_000290-af_000060-minus-0000087", "task": "close double door", "image": [ "images/robocasa_data/close_double_door/episode_34/cam_high/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_34/cam_high/frame_000354.jpg", "images/robocasa_data/close_double_door/episode_35/cam_high/frame_000290.jpg", "images/robocasa_data/close_double_door/episode_35/cam_left_wrist/frame_000290.jpg", "images/robocasa_data/close_double_door/episode_35/cam_right_wrist/frame_000290.jpg", "images/robocasa_data/close_double_door/episode_35/cam_high/frame_000060.jpg", "images/robocasa_data/close_double_door/episode_35/cam_left_wrist/frame_000060.jpg", "images/robocasa_data/close_double_door/episode_35/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -79.3, "hop_absolute_value": 79.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_stove_episode_43-train-bf_000150-af_000000-minus-0000020", "task": "turn off stove", "image": [ "images/robocasa_data/turn_off_stove/episode_40/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_stove/episode_40/cam_high/frame_000164.jpg", "images/robocasa_data/turn_off_stove/episode_43/cam_high/frame_000150.jpg", "images/robocasa_data/turn_off_stove/episode_43/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/turn_off_stove/episode_43/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/turn_off_stove/episode_43/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_stove/episode_43/cam_left_wrist/frame_000000.jpg", "images/robocasa_data/turn_off_stove/episode_43/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_stove_to_counter_episode_16-train-bf_000150-af_000110-minus-0000810", "task": "pick and place stove to counter", "image": [ "images/robocasa_data/pick_and_place_stove_to_counter/episode_9/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_9/cam_high/frame_000288.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_high/frame_000150.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_high/frame_000110.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_16/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place stove to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place stove to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -26.7, "hop_absolute_value": 26.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_stove_to_counter_episode_23-train-bf_000210-af_000140-minus-0000653", "task": "pick and place stove to counter", "image": [ "images/robocasa_data/pick_and_place_stove_to_counter/episode_7/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_7/cam_high/frame_000264.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_23/cam_high/frame_000210.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_23/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_23/cam_right_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_23/cam_high/frame_000140.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_23/cam_left_wrist/frame_000140.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_23/cam_right_wrist/frame_000140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place stove to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place stove to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_single_door_episode_22-train-bf_000270-af_000160-minus-0000209", "task": "close single door", "image": [ "images/robocasa_data/close_single_door/episode_18/cam_high/frame_000000.jpg", "images/robocasa_data/close_single_door/episode_18/cam_high/frame_000307.jpg", "images/robocasa_data/close_single_door/episode_22/cam_high/frame_000270.jpg", "images/robocasa_data/close_single_door/episode_22/cam_left_wrist/frame_000270.jpg", "images/robocasa_data/close_single_door/episode_22/cam_right_wrist/frame_000270.jpg", "images/robocasa_data/close_single_door/episode_22/cam_high/frame_000160.jpg", "images/robocasa_data/close_single_door/episode_22/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/close_single_door/episode_22/cam_right_wrist/frame_000160.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.7, "hop_absolute_value": 40.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_sink_episode_25-train-bf_000150-af_000010-minus-0000013", "task": "pick and place counter to sink", "image": [ "images/robocasa_data/pick_and_place_counter_to_sink/episode_14/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_14/cam_high/frame_000593.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_25/cam_high/frame_000150.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_25/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_25/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_25/cam_high/frame_000010.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_25/cam_left_wrist/frame_000010.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_25/cam_right_wrist/frame_000010.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -93.3, "hop_absolute_value": 93.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_25-train-bf_000600-af_000550-minus-0000795", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_34/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_34/cam_high/frame_000819.jpg", "images/robocasa_data/open_double_door/episode_25/cam_high/frame_000600.jpg", "images/robocasa_data/open_double_door/episode_25/cam_left_wrist/frame_000600.jpg", "images/robocasa_data/open_double_door/episode_25/cam_right_wrist/frame_000600.jpg", "images/robocasa_data/open_double_door/episode_25/cam_high/frame_000550.jpg", "images/robocasa_data/open_double_door/episode_25/cam_left_wrist/frame_000550.jpg", "images/robocasa_data/open_double_door/episode_25/cam_right_wrist/frame_000550.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -8.3, "hop_absolute_value": 8.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_microwave_episode_3-train-bf_000160-af_000140-minus-0000249", "task": "turn off microwave", "image": [ "images/robocasa_data/turn_off_microwave/episode_1/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_microwave/episode_1/cam_high/frame_000297.jpg", "images/robocasa_data/turn_off_microwave/episode_3/cam_high/frame_000160.jpg", "images/robocasa_data/turn_off_microwave/episode_3/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/turn_off_microwave/episode_3/cam_right_wrist/frame_000160.jpg", "images/robocasa_data/turn_off_microwave/episode_3/cam_high/frame_000140.jpg", "images/robocasa_data/turn_off_microwave/episode_3/cam_left_wrist/frame_000140.jpg", "images/robocasa_data/turn_off_microwave/episode_3/cam_right_wrist/frame_000140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off microwave`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off microwave`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_4-train-bf_000230-af_000000-minus-0000018", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_4/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_4/cam_high/frame_000423.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_4/cam_high/frame_000230.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_4/cam_left_wrist/frame_000230.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_4/cam_right_wrist/frame_000230.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_4/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_4/cam_left_wrist/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_4/cam_right_wrist/frame_000000.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_single_door_episode_0-train-bf_000130-af_000070-minus-0000116", "task": "close single door", "image": [ "images/robocasa_data/close_single_door/episode_31/cam_high/frame_000000.jpg", "images/robocasa_data/close_single_door/episode_31/cam_high/frame_000326.jpg", "images/robocasa_data/close_single_door/episode_0/cam_high/frame_000130.jpg", "images/robocasa_data/close_single_door/episode_0/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/close_single_door/episode_0/cam_right_wrist/frame_000130.jpg", "images/robocasa_data/close_single_door/episode_0/cam_high/frame_000070.jpg", "images/robocasa_data/close_single_door/episode_0/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/close_single_door/episode_0/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -46.2, "hop_absolute_value": 46.2, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_stove_episode_36-train-bf_000310-af_000200-minus-0000245", "task": "pick and place counter to stove", "image": [ "images/robocasa_data/pick_and_place_counter_to_stove/episode_6/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_6/cam_high/frame_000256.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_36/cam_high/frame_000310.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_36/cam_left_wrist/frame_000310.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_36/cam_right_wrist/frame_000310.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_36/cam_high/frame_000200.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_36/cam_left_wrist/frame_000200.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_36/cam_right_wrist/frame_000200.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -35.5, "hop_absolute_value": 35.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_drawer_episode_2472-train-bf_000190-af_000030-minus-0000040", "task": "open drawer", "image": [ "images/robocasa_data/open_drawer/episode_2472/cam_high/frame_000000.jpg", "images/robocasa_data/open_drawer/episode_2472/cam_high/frame_000219.jpg", "images/robocasa_data/open_drawer/episode_2472/cam_high/frame_000190.jpg", "images/robocasa_data/open_drawer/episode_2472/cam_left_wrist/frame_000190.jpg", "images/robocasa_data/open_drawer/episode_2472/cam_right_wrist/frame_000190.jpg", "images/robocasa_data/open_drawer/episode_2472/cam_high/frame_000030.jpg", "images/robocasa_data/open_drawer/episode_2472/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/open_drawer/episode_2472/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -84.2, "hop_absolute_value": 84.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_0-train-bf_000090-af_000070-minus-0000703", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_high/frame_000339.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_high/frame_000090.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_right_wrist/frame_000090.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_high/frame_000070.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_0/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -22.2, "hop_absolute_value": 22.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_stove_episode_12-train-bf_000120-af_000080-minus-0000108", "task": "turn on stove", "image": [ "images/robocasa_data/turn_on_stove/episode_3/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_stove/episode_3/cam_high/frame_000206.jpg", "images/robocasa_data/turn_on_stove/episode_12/cam_high/frame_000120.jpg", "images/robocasa_data/turn_on_stove/episode_12/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/turn_on_stove/episode_12/cam_right_wrist/frame_000120.jpg", "images/robocasa_data/turn_on_stove/episode_12/cam_high/frame_000080.jpg", "images/robocasa_data/turn_on_stove/episode_12/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/turn_on_stove/episode_12/cam_right_wrist/frame_000080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_double_door_episode_22-train-bf_000580-af_000450-minus-0000266", "task": "open double door", "image": [ "images/robocasa_data/open_double_door/episode_25/cam_high/frame_000000.jpg", "images/robocasa_data/open_double_door/episode_25/cam_high/frame_000675.jpg", "images/robocasa_data/open_double_door/episode_22/cam_high/frame_000580.jpg", "images/robocasa_data/open_double_door/episode_22/cam_left_wrist/frame_000580.jpg", "images/robocasa_data/open_double_door/episode_22/cam_right_wrist/frame_000580.jpg", "images/robocasa_data/open_double_door/episode_22/cam_high/frame_000450.jpg", "images/robocasa_data/open_double_door/episode_22/cam_left_wrist/frame_000450.jpg", "images/robocasa_data/open_double_door/episode_22/cam_right_wrist/frame_000450.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -22.4, "hop_absolute_value": 22.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_6-train-bf_000140-af_000070-minus-0000906", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_31/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_31/cam_high/frame_000452.jpg", "images/robocasa_data/coffee_serve_mug/episode_6/cam_high/frame_000140.jpg", "images/robocasa_data/coffee_serve_mug/episode_6/cam_left_wrist/frame_000140.jpg", "images/robocasa_data/coffee_serve_mug/episode_6/cam_right_wrist/frame_000140.jpg", "images/robocasa_data/coffee_serve_mug/episode_6/cam_high/frame_000070.jpg", "images/robocasa_data/coffee_serve_mug/episode_6/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/coffee_serve_mug/episode_6/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_setup_mug_episode_12-train-bf_000080-af_000050-minus-0000176", "task": "coffee setup mug", "image": [ "images/robocasa_data/coffee_setup_mug/episode_12/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_12/cam_high/frame_000254.jpg", "images/robocasa_data/coffee_setup_mug/episode_12/cam_high/frame_000080.jpg", "images/robocasa_data/coffee_setup_mug/episode_12/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/coffee_setup_mug/episode_12/cam_right_wrist/frame_000080.jpg", "images/robocasa_data/coffee_setup_mug/episode_12/cam_high/frame_000050.jpg", "images/robocasa_data/coffee_setup_mug/episode_12/cam_left_wrist/frame_000050.jpg", "images/robocasa_data/coffee_setup_mug/episode_12/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee setup mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee setup mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -37.5, "hop_absolute_value": 37.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_drawer_episode_378-train-bf_000080-af_000050-minus-0000136", "task": "close drawer", "image": [ "images/robocasa_data/close_drawer/episode_377/cam_high/frame_000000.jpg", "images/robocasa_data/close_drawer/episode_377/cam_high/frame_000212.jpg", "images/robocasa_data/close_drawer/episode_378/cam_high/frame_000080.jpg", "images/robocasa_data/close_drawer/episode_378/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/close_drawer/episode_378/cam_right_wrist/frame_000080.jpg", "images/robocasa_data/close_drawer/episode_378/cam_high/frame_000050.jpg", "images/robocasa_data/close_drawer/episode_378/cam_left_wrist/frame_000050.jpg", "images/robocasa_data/close_drawer/episode_378/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -37.5, "hop_absolute_value": 37.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_single_door_episode_15-train-bf_000140-af_000020-minus-0000043", "task": "close single door", "image": [ "images/robocasa_data/close_single_door/episode_15/cam_high/frame_000000.jpg", "images/robocasa_data/close_single_door/episode_15/cam_high/frame_000294.jpg", "images/robocasa_data/close_single_door/episode_15/cam_high/frame_000140.jpg", "images/robocasa_data/close_single_door/episode_15/cam_left_wrist/frame_000140.jpg", "images/robocasa_data/close_single_door/episode_15/cam_right_wrist/frame_000140.jpg", "images/robocasa_data/close_single_door/episode_15/cam_high/frame_000020.jpg", "images/robocasa_data/close_single_door/episode_15/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/close_single_door/episode_15/cam_right_wrist/frame_000020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -85.7, "hop_absolute_value": 85.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_8-train-bf_000160-af_000120-minus-0000600", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_34/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_34/cam_high/frame_000318.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_8/cam_high/frame_000160.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_8/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_8/cam_right_wrist/frame_000160.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_8/cam_high/frame_000120.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_8/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_8/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_setup_mug_episode_49-train-bf_000130-af_000110-minus-0000560", "task": "coffee setup mug", "image": [ "images/robocasa_data/coffee_setup_mug/episode_50/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_50/cam_high/frame_000339.jpg", "images/robocasa_data/coffee_setup_mug/episode_49/cam_high/frame_000130.jpg", "images/robocasa_data/coffee_setup_mug/episode_49/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/coffee_setup_mug/episode_49/cam_right_wrist/frame_000130.jpg", "images/robocasa_data/coffee_setup_mug/episode_49/cam_high/frame_000110.jpg", "images/robocasa_data/coffee_setup_mug/episode_49/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/coffee_setup_mug/episode_49/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee setup mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee setup mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -15.4, "hop_absolute_value": 15.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_4-train-bf_000340-af_000140-minus-0000185", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_9/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_9/cam_high/frame_000360.jpg", "images/robocasa_data/coffee_serve_mug/episode_4/cam_high/frame_000340.jpg", "images/robocasa_data/coffee_serve_mug/episode_4/cam_left_wrist/frame_000340.jpg", "images/robocasa_data/coffee_serve_mug/episode_4/cam_right_wrist/frame_000340.jpg", "images/robocasa_data/coffee_serve_mug/episode_4/cam_high/frame_000140.jpg", "images/robocasa_data/coffee_serve_mug/episode_4/cam_left_wrist/frame_000140.jpg", "images/robocasa_data/coffee_serve_mug/episode_4/cam_right_wrist/frame_000140.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -58.8, "hop_absolute_value": 58.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_stove_episode_25-train-bf_000220-af_000030-minus-0000053", "task": "turn off stove", "image": [ "images/robocasa_data/turn_off_stove/episode_12/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_stove/episode_12/cam_high/frame_000143.jpg", "images/robocasa_data/turn_off_stove/episode_25/cam_high/frame_000220.jpg", "images/robocasa_data/turn_off_stove/episode_25/cam_left_wrist/frame_000220.jpg", "images/robocasa_data/turn_off_stove/episode_25/cam_right_wrist/frame_000220.jpg", "images/robocasa_data/turn_off_stove/episode_25/cam_high/frame_000030.jpg", "images/robocasa_data/turn_off_stove/episode_25/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/turn_off_stove/episode_25/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -86.4, "hop_absolute_value": 86.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_stove_episode_3-train-bf_000280-af_000170-minus-0000771", "task": "pick and place counter to stove", "image": [ "images/robocasa_data/pick_and_place_counter_to_stove/episode_3/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_3/cam_high/frame_000319.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_3/cam_high/frame_000280.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_3/cam_left_wrist/frame_000280.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_3/cam_right_wrist/frame_000280.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_3/cam_high/frame_000170.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_3/cam_left_wrist/frame_000170.jpg", "images/robocasa_data/pick_and_place_counter_to_stove/episode_3/cam_right_wrist/frame_000170.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -39.3, "hop_absolute_value": 39.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_double_door_episode_35-train-bf_000320-af_000230-minus-0000877", "task": "close double door", "image": [ "images/robocasa_data/close_double_door/episode_27/cam_high/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_27/cam_high/frame_000365.jpg", "images/robocasa_data/close_double_door/episode_35/cam_high/frame_000320.jpg", "images/robocasa_data/close_double_door/episode_35/cam_left_wrist/frame_000320.jpg", "images/robocasa_data/close_double_door/episode_35/cam_right_wrist/frame_000320.jpg", "images/robocasa_data/close_double_door/episode_35/cam_high/frame_000230.jpg", "images/robocasa_data/close_double_door/episode_35/cam_left_wrist/frame_000230.jpg", "images/robocasa_data/close_double_door/episode_35/cam_right_wrist/frame_000230.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -28.1, "hop_absolute_value": 28.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_sink_faucet_episode_5-train-bf_000080-af_000020-minus-0000039", "task": "turn on sink faucet", "image": [ "images/robocasa_data/turn_on_sink_faucet/episode_9/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_9/cam_high/frame_000334.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_5/cam_high/frame_000080.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_5/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_5/cam_right_wrist/frame_000080.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_5/cam_high/frame_000020.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_5/cam_left_wrist/frame_000020.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_5/cam_right_wrist/frame_000020.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_double_door_episode_40-train-bf_000310-af_000210-minus-0000913", "task": "close double door", "image": [ "images/robocasa_data/close_double_door/episode_5/cam_high/frame_000000.jpg", "images/robocasa_data/close_double_door/episode_5/cam_high/frame_000382.jpg", "images/robocasa_data/close_double_door/episode_40/cam_high/frame_000310.jpg", "images/robocasa_data/close_double_door/episode_40/cam_left_wrist/frame_000310.jpg", "images/robocasa_data/close_double_door/episode_40/cam_right_wrist/frame_000310.jpg", "images/robocasa_data/close_double_door/episode_40/cam_high/frame_000210.jpg", "images/robocasa_data/close_double_door/episode_40/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/close_double_door/episode_40/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close double door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close double door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -32.3, "hop_absolute_value": 32.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_19-train-bf_000150-af_000060-minus-0000162", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_26/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_26/cam_high/frame_000311.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_19/cam_high/frame_000150.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_19/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_19/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_19/cam_high/frame_000060.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_19/cam_left_wrist/frame_000060.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_19/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_sink_episode_35-train-bf_000330-af_000230-minus-0000245", "task": "pick and place counter to sink", "image": [ "images/robocasa_data/pick_and_place_counter_to_sink/episode_42/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_42/cam_high/frame_000461.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_35/cam_high/frame_000330.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_35/cam_left_wrist/frame_000330.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_35/cam_right_wrist/frame_000330.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_35/cam_high/frame_000230.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_35/cam_left_wrist/frame_000230.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_35/cam_right_wrist/frame_000230.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -30.3, "hop_absolute_value": 30.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_microwave_to_counter_episode_5-train-bf_000250-af_000200-minus-0000234", "task": "pick and place microwave to counter", "image": [ "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_15/cam_high/frame_000352.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_5/cam_high/frame_000250.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_5/cam_left_wrist/frame_000250.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_5/cam_right_wrist/frame_000250.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_5/cam_high/frame_000200.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_5/cam_left_wrist/frame_000200.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_5/cam_right_wrist/frame_000200.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place microwave to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place microwave to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_stove_episode_35-train-bf_000150-af_000040-minus-0000109", "task": "turn on stove", "image": [ "images/robocasa_data/turn_on_stove/episode_37/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_stove/episode_37/cam_high/frame_000222.jpg", "images/robocasa_data/turn_on_stove/episode_35/cam_high/frame_000150.jpg", "images/robocasa_data/turn_on_stove/episode_35/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/turn_on_stove/episode_35/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/turn_on_stove/episode_35/cam_high/frame_000040.jpg", "images/robocasa_data/turn_on_stove/episode_35/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/turn_on_stove/episode_35/cam_right_wrist/frame_000040.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.3, "hop_absolute_value": 73.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_sink_faucet_episode_18-train-bf_000160-af_000050-minus-0000100", "task": "turn on sink faucet", "image": [ "images/robocasa_data/turn_on_sink_faucet/episode_27/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_27/cam_high/frame_000121.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_18/cam_high/frame_000160.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_18/cam_left_wrist/frame_000160.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_18/cam_right_wrist/frame_000160.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_18/cam_high/frame_000050.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_18/cam_left_wrist/frame_000050.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_18/cam_right_wrist/frame_000050.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -68.8, "hop_absolute_value": 68.8, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_counter_to_sink_episode_48-train-bf_000430-af_000030-minus-0000035", "task": "pick and place counter to sink", "image": [ "images/robocasa_data/pick_and_place_counter_to_sink/episode_32/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_32/cam_high/frame_000399.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_48/cam_high/frame_000430.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_48/cam_left_wrist/frame_000430.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_48/cam_right_wrist/frame_000430.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_48/cam_high/frame_000030.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_48/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/pick_and_place_counter_to_sink/episode_48/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place counter to sink`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place counter to sink`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -93.0, "hop_absolute_value": 93.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_drawer_episode_2455-train-bf_000170-af_000120-minus-0000485", "task": "open drawer", "image": [ "images/robocasa_data/open_drawer/episode_2459/cam_high/frame_000000.jpg", "images/robocasa_data/open_drawer/episode_2459/cam_high/frame_000241.jpg", "images/robocasa_data/open_drawer/episode_2455/cam_high/frame_000170.jpg", "images/robocasa_data/open_drawer/episode_2455/cam_left_wrist/frame_000170.jpg", "images/robocasa_data/open_drawer/episode_2455/cam_right_wrist/frame_000170.jpg", "images/robocasa_data/open_drawer/episode_2455/cam_high/frame_000120.jpg", "images/robocasa_data/open_drawer/episode_2455/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/open_drawer/episode_2455/cam_right_wrist/frame_000120.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -29.4, "hop_absolute_value": 29.4, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_close_single_door_episode_15-train-bf_000230-af_000060-minus-0000114", "task": "close single door", "image": [ "images/robocasa_data/close_single_door/episode_32/cam_high/frame_000000.jpg", "images/robocasa_data/close_single_door/episode_32/cam_high/frame_000271.jpg", "images/robocasa_data/close_single_door/episode_15/cam_high/frame_000230.jpg", "images/robocasa_data/close_single_door/episode_15/cam_left_wrist/frame_000230.jpg", "images/robocasa_data/close_single_door/episode_15/cam_right_wrist/frame_000230.jpg", "images/robocasa_data/close_single_door/episode_15/cam_high/frame_000060.jpg", "images/robocasa_data/close_single_door/episode_15/cam_left_wrist/frame_000060.jpg", "images/robocasa_data/close_single_door/episode_15/cam_right_wrist/frame_000060.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`close single door`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `close single door`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.9, "hop_absolute_value": 73.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_setup_mug_episode_50-train-bf_000090-af_000080-minus-0000567", "task": "coffee setup mug", "image": [ "images/robocasa_data/coffee_setup_mug/episode_50/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_50/cam_high/frame_000339.jpg", "images/robocasa_data/coffee_setup_mug/episode_50/cam_high/frame_000090.jpg", "images/robocasa_data/coffee_setup_mug/episode_50/cam_left_wrist/frame_000090.jpg", "images/robocasa_data/coffee_setup_mug/episode_50/cam_right_wrist/frame_000090.jpg", "images/robocasa_data/coffee_setup_mug/episode_50/cam_high/frame_000080.jpg", "images/robocasa_data/coffee_setup_mug/episode_50/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/coffee_setup_mug/episode_50/cam_right_wrist/frame_000080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee setup mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee setup mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_sink_faucet_episode_0-train-bf_000120-af_000080-minus-0000080", "task": "turn off sink faucet", "image": [ "images/robocasa_data/turn_off_sink_faucet/episode_14/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_14/cam_high/frame_000154.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_0/cam_high/frame_000120.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_0/cam_left_wrist/frame_000120.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_0/cam_right_wrist/frame_000120.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_0/cam_high/frame_000080.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_0/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_0/cam_right_wrist/frame_000080.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_stove_to_counter_episode_0-train-bf_000100-af_000040-minus-0000127", "task": "pick and place stove to counter", "image": [ "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_high/frame_000334.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_0/cam_high/frame_000100.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_0/cam_left_wrist/frame_000100.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_0/cam_right_wrist/frame_000100.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_0/cam_high/frame_000040.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_0/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_0/cam_right_wrist/frame_000040.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place stove to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place stove to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_setup_mug_episode_1-train-bf_000170-af_000130-minus-0000518", "task": "coffee setup mug", "image": [ "images/robocasa_data/coffee_setup_mug/episode_42/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_42/cam_high/frame_000370.jpg", "images/robocasa_data/coffee_setup_mug/episode_1/cam_high/frame_000170.jpg", "images/robocasa_data/coffee_setup_mug/episode_1/cam_left_wrist/frame_000170.jpg", "images/robocasa_data/coffee_setup_mug/episode_1/cam_right_wrist/frame_000170.jpg", "images/robocasa_data/coffee_setup_mug/episode_1/cam_high/frame_000130.jpg", "images/robocasa_data/coffee_setup_mug/episode_1/cam_left_wrist/frame_000130.jpg", "images/robocasa_data/coffee_setup_mug/episode_1/cam_right_wrist/frame_000130.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee setup mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee setup mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -23.5, "hop_absolute_value": 23.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_microwave_to_counter_episode_7-train-bf_000260-af_000150-minus-0000212", "task": "pick and place microwave to counter", "image": [ "images/robocasa_data/pick_and_place_microwave_to_counter/episode_7/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_7/cam_high/frame_000324.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_7/cam_high/frame_000260.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_7/cam_left_wrist/frame_000260.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_7/cam_right_wrist/frame_000260.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_7/cam_high/frame_000150.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_7/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_7/cam_right_wrist/frame_000150.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place microwave to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place microwave to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -42.3, "hop_absolute_value": 42.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_setup_mug_episode_7-train-bf_000300-af_000200-minus-0000240", "task": "coffee setup mug", "image": [ "images/robocasa_data/coffee_setup_mug/episode_7/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_high/frame_000364.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_high/frame_000300.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_left_wrist/frame_000300.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_right_wrist/frame_000300.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_high/frame_000200.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_left_wrist/frame_000200.jpg", "images/robocasa_data/coffee_setup_mug/episode_7/cam_right_wrist/frame_000200.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee setup mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee setup mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_cab_to_counter_episode_20-train-bf_000240-af_000030-minus-0000051", "task": "pick and place cab to counter", "image": [ "images/robocasa_data/pick_and_place_cab_to_counter/episode_4/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_4/cam_high/frame_000423.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_20/cam_high/frame_000240.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_20/cam_left_wrist/frame_000240.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_20/cam_right_wrist/frame_000240.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_20/cam_high/frame_000030.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_20/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/pick_and_place_cab_to_counter/episode_20/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place cab to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place cab to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_microwave_to_counter_episode_20-train-bf_000270-af_000230-minus-0000623", "task": "pick and place microwave to counter", "image": [ "images/robocasa_data/pick_and_place_microwave_to_counter/episode_12/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_12/cam_high/frame_000330.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_20/cam_high/frame_000270.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_20/cam_left_wrist/frame_000270.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_20/cam_right_wrist/frame_000270.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_20/cam_high/frame_000230.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_20/cam_left_wrist/frame_000230.jpg", "images/robocasa_data/pick_and_place_microwave_to_counter/episode_20/cam_right_wrist/frame_000230.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place microwave to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place microwave to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -14.8, "hop_absolute_value": 14.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_stove_to_counter_episode_20-train-bf_000240-af_000210-minus-0000257", "task": "pick and place stove to counter", "image": [ "images/robocasa_data/pick_and_place_stove_to_counter/episode_30/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_30/cam_high/frame_000274.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_high/frame_000240.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_left_wrist/frame_000240.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_right_wrist/frame_000240.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_high/frame_000210.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/pick_and_place_stove_to_counter/episode_20/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place stove to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place stove to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_stove_episode_31-train-bf_000110-af_000010-minus-0000017", "task": "turn off stove", "image": [ "images/robocasa_data/turn_off_stove/episode_56/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_stove/episode_56/cam_high/frame_000186.jpg", "images/robocasa_data/turn_off_stove/episode_31/cam_high/frame_000110.jpg", "images/robocasa_data/turn_off_stove/episode_31/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/turn_off_stove/episode_31/cam_right_wrist/frame_000110.jpg", "images/robocasa_data/turn_off_stove/episode_31/cam_high/frame_000010.jpg", "images/robocasa_data/turn_off_stove/episode_31/cam_left_wrist/frame_000010.jpg", "images/robocasa_data/turn_off_stove/episode_31/cam_right_wrist/frame_000010.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -90.9, "hop_absolute_value": 90.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_sink_faucet_episode_7-train-bf_000070-af_000030-minus-0000049", "task": "turn on sink faucet", "image": [ "images/robocasa_data/turn_on_sink_faucet/episode_26/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_26/cam_high/frame_000150.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_7/cam_high/frame_000070.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_7/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_7/cam_right_wrist/frame_000070.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_7/cam_high/frame_000030.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_7/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_7/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -57.1, "hop_absolute_value": 57.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_pick_and_place_sink_to_counter_episode_12-train-bf_000150-af_000040-minus-0000083", "task": "pick and place sink to counter", "image": [ "images/robocasa_data/pick_and_place_sink_to_counter/episode_34/cam_high/frame_000000.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_34/cam_high/frame_000277.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_12/cam_high/frame_000150.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_12/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_12/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_12/cam_high/frame_000040.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_12/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/pick_and_place_sink_to_counter/episode_12/cam_right_wrist/frame_000040.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`pick and place sink to counter`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `pick and place sink to counter`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.3, "hop_absolute_value": 73.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_sink_faucet_episode_23-train-bf_000040-af_000030-minus-0000194", "task": "turn off sink faucet", "image": [ "images/robocasa_data/turn_off_sink_faucet/episode_23/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_23/cam_high/frame_000236.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_23/cam_high/frame_000040.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_23/cam_left_wrist/frame_000040.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_23/cam_right_wrist/frame_000040.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_23/cam_high/frame_000030.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_23/cam_left_wrist/frame_000030.jpg", "images/robocasa_data/turn_off_sink_faucet/episode_23/cam_right_wrist/frame_000030.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_open_drawer_episode_2475-train-bf_000200-af_000110-minus-0000602", "task": "open drawer", "image": [ "images/robocasa_data/open_drawer/episode_2475/cam_high/frame_000000.jpg", "images/robocasa_data/open_drawer/episode_2475/cam_high/frame_000254.jpg", "images/robocasa_data/open_drawer/episode_2475/cam_high/frame_000200.jpg", "images/robocasa_data/open_drawer/episode_2475/cam_left_wrist/frame_000200.jpg", "images/robocasa_data/open_drawer/episode_2475/cam_right_wrist/frame_000200.jpg", "images/robocasa_data/open_drawer/episode_2475/cam_high/frame_000110.jpg", "images/robocasa_data/open_drawer/episode_2475/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/open_drawer/episode_2475/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`open drawer`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `open drawer`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -45.0, "hop_absolute_value": 45.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_0-train-bf_000240-af_000210-minus-0000807", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_31/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_31/cam_high/frame_000452.jpg", "images/robocasa_data/coffee_serve_mug/episode_0/cam_high/frame_000240.jpg", "images/robocasa_data/coffee_serve_mug/episode_0/cam_left_wrist/frame_000240.jpg", "images/robocasa_data/coffee_serve_mug/episode_0/cam_right_wrist/frame_000240.jpg", "images/robocasa_data/coffee_serve_mug/episode_0/cam_high/frame_000210.jpg", "images/robocasa_data/coffee_serve_mug/episode_0/cam_left_wrist/frame_000210.jpg", "images/robocasa_data/coffee_serve_mug/episode_0/cam_right_wrist/frame_000210.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_off_stove_episode_34-train-bf_000150-af_000110-minus-0000119", "task": "turn off stove", "image": [ "images/robocasa_data/turn_off_stove/episode_35/cam_high/frame_000000.jpg", "images/robocasa_data/turn_off_stove/episode_35/cam_high/frame_000140.jpg", "images/robocasa_data/turn_off_stove/episode_34/cam_high/frame_000150.jpg", "images/robocasa_data/turn_off_stove/episode_34/cam_left_wrist/frame_000150.jpg", "images/robocasa_data/turn_off_stove/episode_34/cam_right_wrist/frame_000150.jpg", "images/robocasa_data/turn_off_stove/episode_34/cam_high/frame_000110.jpg", "images/robocasa_data/turn_off_stove/episode_34/cam_left_wrist/frame_000110.jpg", "images/robocasa_data/turn_off_stove/episode_34/cam_right_wrist/frame_000110.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn off stove`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn off stove`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -26.7, "hop_absolute_value": 26.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_turn_on_sink_faucet_episode_3-train-bf_000370-af_000360-minus-0000293", "task": "turn on sink faucet", "image": [ "images/robocasa_data/turn_on_sink_faucet/episode_28/cam_high/frame_000000.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_28/cam_high/frame_000201.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_3/cam_high/frame_000370.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_3/cam_left_wrist/frame_000370.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_3/cam_right_wrist/frame_000370.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_3/cam_high/frame_000360.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_3/cam_left_wrist/frame_000360.jpg", "images/robocasa_data/turn_on_sink_faucet/episode_3/cam_right_wrist/frame_000360.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`turn on sink faucet`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `turn on sink faucet`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.7, "hop_absolute_value": 2.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_press_button_episode_8-train-bf_000080-af_000070-minus-0000098", "task": "coffee press button", "image": [ "images/robocasa_data/coffee_press_button/episode_10/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_press_button/episode_10/cam_high/frame_000138.jpg", "images/robocasa_data/coffee_press_button/episode_8/cam_high/frame_000080.jpg", "images/robocasa_data/coffee_press_button/episode_8/cam_left_wrist/frame_000080.jpg", "images/robocasa_data/coffee_press_button/episode_8/cam_right_wrist/frame_000080.jpg", "images/robocasa_data/coffee_press_button/episode_8/cam_high/frame_000070.jpg", "images/robocasa_data/coffee_press_button/episode_8/cam_left_wrist/frame_000070.jpg", "images/robocasa_data/coffee_press_button/episode_8/cam_right_wrist/frame_000070.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee press button`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee press button`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_serve_mug_episode_23-train-bf_000230-af_000220-minus-0000572", "task": "coffee serve mug", "image": [ "images/robocasa_data/coffee_serve_mug/episode_23/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_high/frame_000336.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_high/frame_000230.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_left_wrist/frame_000230.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_right_wrist/frame_000230.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_high/frame_000220.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_left_wrist/frame_000220.jpg", "images/robocasa_data/coffee_serve_mug/episode_23/cam_right_wrist/frame_000220.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee serve mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee serve mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.3, "hop_absolute_value": 4.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robocasa_data_coffee_setup_mug_episode_6-train-bf_000180-af_000170-minus-0000482", "task": "coffee setup mug", "image": [ "images/robocasa_data/coffee_setup_mug/episode_53/cam_high/frame_000000.jpg", "images/robocasa_data/coffee_setup_mug/episode_53/cam_high/frame_000268.jpg", "images/robocasa_data/coffee_setup_mug/episode_6/cam_high/frame_000180.jpg", "images/robocasa_data/coffee_setup_mug/episode_6/cam_left_wrist/frame_000180.jpg", "images/robocasa_data/coffee_setup_mug/episode_6/cam_right_wrist/frame_000180.jpg", "images/robocasa_data/coffee_setup_mug/episode_6/cam_high/frame_000170.jpg", "images/robocasa_data/coffee_setup_mug/episode_6/cam_left_wrist/frame_000170.jpg", "images/robocasa_data/coffee_setup_mug/episode_6/cam_right_wrist/frame_000170.jpg" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`coffee setup mug`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `coffee setup mug`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -5.6, "hop_absolute_value": 5.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robocasa_data" }, { "id": "robotwin_rotate_qrcode_episode_014-train-bf_000000-af_000157-plus-0000130", "task": "Lift the payment sign with angled face and orient its QR code toward you with the right arm.", "image": [ "images/robotwin2_agilex_part1/rotate_qrcode_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/rotate_qrcode_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000154.png", "images/robotwin2_agilex_part1/rotate_qrcode_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/rotate_qrcode_aloha-agilex_randomized_500/episode_014/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/rotate_qrcode_aloha-agilex_randomized_500/episode_014/sample_frames/right_camera/000000.png", "images/robotwin2_agilex_part1/rotate_qrcode_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000157.png", "images/robotwin2_agilex_part1/rotate_qrcode_aloha-agilex_randomized_500/episode_014/sample_frames/left_camera/000157.png", "images/robotwin2_agilex_part1/rotate_qrcode_aloha-agilex_randomized_500/episode_014/sample_frames/right_camera/000157.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift the payment sign with angled face and orient its QR code toward you with the right arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift the payment sign with angled face and orient its QR code toward you with the right arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_011-train-bf_000164-af_000191-plus-0000596", "task": "Place red block to the left, green block in the middle, and blue block on the right in a linear fashion.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000461.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000164.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000164.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000164.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000191.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000191.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000191.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place red block to the left, green block in the middle, and blue block on the right in a linear fashion.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place red block to the left, green block in the middle, and blue block on the right in a linear fashion.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.5, "hop_absolute_value": 9.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_003-train-bf_000270-af_000660-plus-0000471", "task": "Open the microwave for heating food by pulling its door with the left arm.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000674.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000270.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_003/sample_frames/left_camera/000270.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_003/sample_frames/right_camera/000270.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000660.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_003/sample_frames/left_camera/000660.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_003/sample_frames/right_camera/000660.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the microwave for heating food by pulling its door with the left arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the microwave for heating food by pulling its door with the left arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 70.3, "hop_absolute_value": 70.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_two_episode_007-train-bf_000187-af_000282-plus-0000331", "task": "Position the small bowl with brown edge carefully on the bowl with brown outer rim.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000328.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_007/sample_frames/head_camera/000187.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_007/sample_frames/left_camera/000187.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_007/sample_frames/right_camera/000187.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_007/sample_frames/head_camera/000282.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_007/sample_frames/left_camera/000282.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_007/sample_frames/right_camera/000282.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Position the small bowl with brown edge carefully on the bowl with brown outer rim.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Position the small bowl with brown edge carefully on the bowl with brown outer rim.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_003-train-bf_000315-af_000645-plus-0000445", "task": "Open the microwave for heating food by pulling its door with the left arm.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000432.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000315.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_003/sample_frames/left_camera/000315.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_003/sample_frames/right_camera/000315.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000645.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_003/sample_frames/left_camera/000645.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_003/sample_frames/right_camera/000645.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the microwave for heating food by pulling its door with the left arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the microwave for heating food by pulling its door with the left arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 64.7, "hop_absolute_value": 64.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_000-train-bf_000147-af_000414-plus-0000464", "task": "Grab the yellow liquid container with green top, the water bottle with textured body, the drink bottle with slim neck, and drop them into the black bin with white liner to the left of the table.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000656.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000147.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_000/sample_frames/left_camera/000147.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_000/sample_frames/right_camera/000147.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000414.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_000/sample_frames/left_camera/000414.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_000/sample_frames/right_camera/000414.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab the yellow liquid container with green top, the water bottle with textured body, the drink bottle with slim neck, and drop them into the black bin with white liner to the left of the table.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab the yellow liquid container with green top, the water bottle with textured body, the drink bottle with slim neck, and drop them into the black bin with white liner to the left of the table.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 69.0, "hop_absolute_value": 69.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_mic_episode_027-train-bf_000098-af_000130-plus-0000158", "task": "Secure the rounded white tip microphone using one arm and transfer it.", "image": [ "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000221.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000098.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000098.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000098.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000130.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000130.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000130.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Secure the rounded white tip microphone using one arm and transfer it.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Secure the rounded white tip microphone using one arm and transfer it.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_block_episode_017-train-bf_000129-af_000157-plus-0000426", "task": "Use the left arm to grasp the red block, transfer it to the right arm, and set it down on the blue pad.", "image": [ "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000285.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000129.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_017/sample_frames/left_camera/000129.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_017/sample_frames/right_camera/000129.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000157.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_017/sample_frames/left_camera/000157.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_017/sample_frames/right_camera/000157.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the left arm to grasp the red block, transfer it to the right arm, and set it down on the blue pad.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the left arm to grasp the red block, transfer it to the right arm, and set it down on the blue pad.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_hanging_mug_episode_018-train-bf_000025-af_000165-plus-0000567", "task": "Grab the mug for hot drinks, rotate it, place it in the middle, and hang it on the rack with angular body.", "image": [ "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000340.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000025.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_018/sample_frames/left_camera/000025.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_018/sample_frames/right_camera/000025.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000165.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_018/sample_frames/left_camera/000165.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_018/sample_frames/right_camera/000165.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab the mug for hot drinks, rotate it, place it in the middle, and hang it on the rack with angular body.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab the mug for hot drinks, rotate it, place it in the middle, and hang it on the rack with angular body.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_three_episode_005-train-bf_000038-af_000154-plus-0000330", "task": "Stack blue block over green block and green block over red block after moving all blocks to the center.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000461.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000038.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_005/sample_frames/left_camera/000038.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_005/sample_frames/right_camera/000038.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000154.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_005/sample_frames/left_camera/000154.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_005/sample_frames/right_camera/000154.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Stack blue block over green block and green block over red block after moving all blocks to the center.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Stack blue block over green block and green block over red block after moving all blocks to the center.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 30.0, "hop_absolute_value": 30.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_020-train-bf_000232-af_000458-plus-0000531", "task": "Using the right arm, the right arm, and the left arm, arrange red block, green block, and blue block from left to right in the given order.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_026/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_026/sample_frames/head_camera/000450.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000232.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/left_camera/000232.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/right_camera/000232.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000458.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/left_camera/000458.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/right_camera/000458.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Using the right arm, the right arm, and the left arm, arrange red block, green block, and blue block from left to right in the given order.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Using the right arm, the right arm, and the left arm, arrange red block, green block, and blue block from left to right in the given order.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_three_episode_027-train-bf_000076-af_000222-plus-0000351", "task": "With the right arm, the left arm, and the right arm, move red block, green block, and blue block centrally and stack green block on red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000472.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000076.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000076.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000076.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000222.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000222.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000222.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`With the right arm, the left arm, and the right arm, move red block, green block, and blue block centrally and stack green block on red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `With the right arm, the left arm, and the right arm, move red block, green block, and blue block centrally and stack green block on red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 40.7, "hop_absolute_value": 40.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_laptop_episode_010-train-bf_000090-af_000195-plus-0000203", "task": "Carefully open the laptop with silver base using the right arm", "image": [ "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000227.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000090.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_010/sample_frames/left_camera/000090.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_010/sample_frames/right_camera/000090.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000195.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_010/sample_frames/left_camera/000195.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_010/sample_frames/right_camera/000195.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Carefully open the laptop with silver base using the right arm`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Carefully open the laptop with silver base using the right arm`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 77.8, "hop_absolute_value": 77.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_mic_episode_002-train-bf_000000-af_000081-plus-0000169", "task": "Pick the white microphone handle black accents and transfer it to the other arm.", "image": [ "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000222.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000000.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000081.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000081.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000081.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the white microphone handle black accents and transfer it to the other arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the white microphone handle black accents and transfer it to the other arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 31.2, "hop_absolute_value": 31.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_024-train-bf_000165-af_000195-plus-0000577", "task": "Utilize the left arm to pull the yellow-panel microwave with smooth sides open.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000490.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000165.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_024/sample_frames/left_camera/000165.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_024/sample_frames/right_camera/000165.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000195.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_024/sample_frames/left_camera/000195.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_024/sample_frames/right_camera/000195.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Utilize the left arm to pull the yellow-panel microwave with smooth sides open.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Utilize the left arm to pull the yellow-panel microwave with smooth sides open.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_007-train-bf_000255-af_000375-plus-0000398", "task": "Access the inside by opening the gray microwave with two control knobs.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000701.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000255.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_007/sample_frames/left_camera/000255.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_007/sample_frames/right_camera/000255.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000375.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_007/sample_frames/left_camera/000375.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_007/sample_frames/right_camera/000375.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Access the inside by opening the gray microwave with two control knobs.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Access the inside by opening the gray microwave with two control knobs.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_three_episode_018-train-bf_000013-af_000154-plus-0000326", "task": "Bring red block, green block, and blue block to the center point and arrange them by stacking blue block atop green block and green block atop red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000461.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000013.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_018/sample_frames/left_camera/000013.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_018/sample_frames/right_camera/000013.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000154.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_018/sample_frames/left_camera/000154.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_018/sample_frames/right_camera/000154.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Bring red block, green block, and blue block to the center point and arrange them by stacking blue block atop green block and green block atop red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Bring red block, green block, and blue block to the center point and arrange them by stacking blue block atop green block and green block atop red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 34.4, "hop_absolute_value": 34.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_can_basket_episode_006-train-bf_000142-af_000197-plus-0000215", "task": "Grab the red-orange text on white can, place it in the basket with arched black grip, and lift the basket with arched black grip upwards.", "image": [ "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000249.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000142.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_006/sample_frames/left_camera/000142.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_006/sample_frames/right_camera/000142.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000197.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_006/sample_frames/left_camera/000197.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_006/sample_frames/right_camera/000197.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab the red-orange text on white can, place it in the basket with arched black grip, and lift the basket with arched black grip upwards.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab the red-orange text on white can, place it in the basket with arched black grip, and lift the basket with arched black grip upwards.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_lift_pot_episode_003-train-bf_000045-af_000090-plus-0000055", "task": "Hold the cylindrical kitchenpot with both arms", "image": [ "images/robotwin2_agilex_part1/lift_pot_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/lift_pot_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000113.png", "images/robotwin2_agilex_part1/lift_pot_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/lift_pot_aloha-agilex_randomized_500/episode_003/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/lift_pot_aloha-agilex_randomized_500/episode_003/sample_frames/right_camera/000045.png", "images/robotwin2_agilex_part1/lift_pot_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000090.png", "images/robotwin2_agilex_part1/lift_pot_aloha-agilex_randomized_500/episode_003/sample_frames/left_camera/000090.png", "images/robotwin2_agilex_part1/lift_pot_aloha-agilex_randomized_500/episode_003/sample_frames/right_camera/000090.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Hold the cylindrical kitchenpot with both arms`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Hold the cylindrical kitchenpot with both arms`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_dump_bin_bigbin_episode_017-train-bf_000120-af_000168-plus-0000149", "task": "Use arm to grab the small trash container, dump contents", "image": [ "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000321.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_017/sample_frames/head_camera/000120.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_017/sample_frames/left_camera/000120.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_017/sample_frames/right_camera/000120.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_017/sample_frames/head_camera/000168.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_017/sample_frames/left_camera/000168.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_017/sample_frames/right_camera/000168.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use arm to grab the small trash container, dump contents`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use arm to grab the small trash container, dump contents`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_phone_stand_episode_003-train-bf_000045-af_000090-plus-0000062", "task": "Move the lightweight phone onto the green phone rack.", "image": [ "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000128.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_003/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_003/sample_frames/right_camera/000045.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000090.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_003/sample_frames/left_camera/000090.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_003/sample_frames/right_camera/000090.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the lightweight phone onto the green phone rack.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the lightweight phone onto the green phone rack.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_dual_shoes_episode_021-train-bf_000060-af_000080-plus-0000138", "task": "Stick both the shoe with black accents in the smooth orange box for shoes, shoe tips pointing left.", "image": [ "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000215.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000060.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_021/sample_frames/left_camera/000060.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_021/sample_frames/right_camera/000060.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000080.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_021/sample_frames/left_camera/000080.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_021/sample_frames/right_camera/000080.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Stick both the shoe with black accents in the smooth orange box for shoes, shoe tips pointing left.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Stick both the shoe with black accents in the smooth orange box for shoes, shoe tips pointing left.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 7.7, "hop_absolute_value": 7.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_021-train-bf_000255-af_000330-plus-0000323", "task": "Grip the microwave with knobs's handle and pull it open.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000674.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000255.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_021/sample_frames/left_camera/000255.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_021/sample_frames/right_camera/000255.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000330.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_021/sample_frames/left_camera/000330.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_021/sample_frames/right_camera/000330.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grip the microwave with knobs's handle and pull it open.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grip the microwave with knobs's handle and pull it open.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 38.5, "hop_absolute_value": 38.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_024-train-bf_000171-af_000248-plus-0000292", "task": "Set red block and green block at the table's center, then stack green block on red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000310.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000171.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_024/sample_frames/left_camera/000171.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_024/sample_frames/right_camera/000171.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000248.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_024/sample_frames/left_camera/000248.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_024/sample_frames/right_camera/000248.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Set red block and green block at the table's center, then stack green block on red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Set red block and green block at the table's center, then stack green block on red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 55.6, "hop_absolute_value": 55.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_hanging_mug_episode_007-train-bf_000038-af_000180-plus-0000326", "task": "Take the hand-sized mug with blue artwork from the table, spin it, place it in the center, then hang it on the smooth rack.", "image": [ "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000339.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000038.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_007/sample_frames/left_camera/000038.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_007/sample_frames/right_camera/000038.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000180.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_007/sample_frames/left_camera/000180.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_007/sample_frames/right_camera/000180.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the hand-sized mug with blue artwork from the table, spin it, place it in the center, then hang it on the smooth rack.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the hand-sized mug with blue artwork from the table, spin it, place it in the center, then hang it on the smooth rack.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 52.4, "hop_absolute_value": 52.4, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_cans_plasticbox_episode_009-train-bf_000030-af_000231-plus-0000327", "task": "Utilize the arms to deposit both the beverage can with printed label and the green red metallic soda can into the plasticbox with smooth plastic texture.", "image": [ "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000285.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000030.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_009/sample_frames/left_camera/000030.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_009/sample_frames/right_camera/000030.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000231.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_009/sample_frames/left_camera/000231.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_009/sample_frames/right_camera/000231.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Utilize the arms to deposit both the beverage can with printed label and the green red metallic soda can into the plasticbox with smooth plastic texture.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Utilize the arms to deposit both the beverage can with printed label and the green red metallic soda can into the plasticbox with smooth plastic texture.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 77.8, "hop_absolute_value": 77.8, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_001-train-bf_000190-af_000307-plus-0000376", "task": "Center red block using the right arm, then stack green block with the left arm on top.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000309.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_001/sample_frames/head_camera/000190.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_001/sample_frames/left_camera/000190.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_001/sample_frames/right_camera/000190.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_001/sample_frames/head_camera/000307.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_001/sample_frames/left_camera/000307.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_001/sample_frames/right_camera/000307.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Center red block using the right arm, then stack green block with the left arm on top.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Center red block using the right arm, then stack green block with the left arm on top.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_block_episode_007-train-bf_000043-af_000071-plus-0000208", "task": "Pick up the red block using the left arm, give it to the right arm, and place it on the blue pad.", "image": [ "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000278.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000043.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_007/sample_frames/left_camera/000043.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_007/sample_frames/right_camera/000043.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000071.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_007/sample_frames/left_camera/000071.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_007/sample_frames/right_camera/000071.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the red block using the left arm, give it to the right arm, and place it on the blue pad.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the red block using the left arm, give it to the right arm, and place it on the blue pad.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_009-train-bf_000076-af_000140-plus-0000493", "task": "Use the left arm to place red block in the center; stack green block on it with the right arm.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000306.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000076.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_009/sample_frames/left_camera/000076.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_009/sample_frames/right_camera/000076.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000140.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_009/sample_frames/left_camera/000140.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_009/sample_frames/right_camera/000140.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the left arm to place red block in the center; stack green block on it with the right arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the left arm to place red block in the center; stack green block on it with the right arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 31.2, "hop_absolute_value": 31.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_click_bell_episode_002-train-bf_000000-af_000045-plus-0000029", "task": "Click the white bell's using the left arm", "image": [ "images/robotwin2_agilex_part1/click_bell_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/click_bell_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000078.png", "images/robotwin2_agilex_part1/click_bell_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/click_bell_aloha-agilex_clean_50/episode_002/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/click_bell_aloha-agilex_clean_50/episode_002/sample_frames/right_camera/000000.png", "images/robotwin2_agilex_part1/click_bell_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/click_bell_aloha-agilex_clean_50/episode_002/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/click_bell_aloha-agilex_clean_50/episode_002/sample_frames/right_camera/000045.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Click the white bell's using the left arm`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Click the white bell's using the left arm`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_shake_bottle_episode_006-train-bf_000015-af_000225-plus-0000275", "task": "Pick up the handheld medium-sized liquid bottle and shake it.", "image": [ "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000242.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000015.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_006/sample_frames/left_camera/000015.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_006/sample_frames/right_camera/000015.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000225.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_006/sample_frames/left_camera/000225.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_006/sample_frames/right_camera/000225.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the handheld medium-sized liquid bottle and shake it.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the handheld medium-sized liquid bottle and shake it.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_020-train-bf_000177-af_000444-plus-0000544", "task": "Using the right arm, the right arm, and the left arm, arrange red block, green block, and blue block from left to right in the given order.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000459.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000177.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/left_camera/000177.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/right_camera/000177.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000444.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/left_camera/000444.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/right_camera/000444.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Using the right arm, the right arm, and the left arm, arrange red block, green block, and blue block from left to right in the given order.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Using the right arm, the right arm, and the left arm, arrange red block, green block, and blue block from left to right in the given order.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 95.0, "hop_absolute_value": 95.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_can_basket_episode_004-train-bf_000036-af_000249-plus-0000314", "task": "Lift the red beverage can with curved top, set it in the light brown oval basket, and elevate the light brown oval basket.", "image": [ "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000244.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000036.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/left_camera/000036.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/right_camera/000036.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000249.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/left_camera/000249.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/right_camera/000249.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift the red beverage can with curved top, set it in the light brown oval basket, and elevate the light brown oval basket.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift the red beverage can with curved top, set it in the light brown oval basket, and elevate the light brown oval basket.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_fan_episode_019-train-bf_000030-af_000147-plus-0000108", "task": "Use the right arm to grab the fan with rectangular base, then align it on the Silver mat facing the robot.", "image": [ "images/robotwin2_agilex_part1/place_fan_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_fan_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000153.png", "images/robotwin2_agilex_part1/place_fan_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000030.png", "images/robotwin2_agilex_part1/place_fan_aloha-agilex_randomized_500/episode_019/sample_frames/left_camera/000030.png", "images/robotwin2_agilex_part1/place_fan_aloha-agilex_randomized_500/episode_019/sample_frames/right_camera/000030.png", "images/robotwin2_agilex_part1/place_fan_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000147.png", "images/robotwin2_agilex_part1/place_fan_aloha-agilex_randomized_500/episode_019/sample_frames/left_camera/000147.png", "images/robotwin2_agilex_part1/place_fan_aloha-agilex_randomized_500/episode_019/sample_frames/right_camera/000147.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the right arm to grab the fan with rectangular base, then align it on the Silver mat facing the robot.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the right arm to grab the fan with rectangular base, then align it on the Silver mat facing the robot.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_004-train-bf_000000-af_000241-plus-0000430", "task": "Take the bottle made of smooth plastic, drop it in the plastic dustbin, then move the medium-sized water bottle and the red bottle to the plastic dustbin.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000519.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_004/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_004/sample_frames/right_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000241.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_004/sample_frames/left_camera/000241.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_004/sample_frames/right_camera/000241.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the bottle made of smooth plastic, drop it in the plastic dustbin, then move the medium-sized water bottle and the red bottle to the plastic dustbin.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the bottle made of smooth plastic, drop it in the plastic dustbin, then move the medium-sized water bottle and the red bottle to the plastic dustbin.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 63.9, "hop_absolute_value": 63.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_006-train-bf_000093-af_000551-plus-0000539", "task": "Begin with the medium yellow cylindrical bottle, move it to the black garbage bin, then continue with the bottle with narrow neck and the red-labeled soda bottle.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000664.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000093.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000093.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000093.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000551.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000551.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000551.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Begin with the medium yellow cylindrical bottle, move it to the black garbage bin, then continue with the bottle with narrow neck and the red-labeled soda bottle.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Begin with the medium yellow cylindrical bottle, move it to the black garbage bin, then continue with the bottle with narrow neck and the red-labeled soda bottle.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 88.6, "hop_absolute_value": 88.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_shake_bottle_horizontally_episode_001-train-bf_000015-af_000165-plus-0000260", "task": "Lift the black cylindrical bottle with ridges and move it side-to-side.", "image": [ "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000279.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_001/sample_frames/head_camera/000015.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_001/sample_frames/left_camera/000015.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_001/sample_frames/right_camera/000015.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_001/sample_frames/head_camera/000165.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_001/sample_frames/left_camera/000165.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_001/sample_frames/right_camera/000165.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift the black cylindrical bottle with ridges and move it side-to-side.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift the black cylindrical bottle with ridges and move it side-to-side.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 55.6, "hop_absolute_value": 55.6, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_three_episode_016-train-bf_000025-af_000102-plus-0000698", "task": "Align bowls the blue bowl in descending size, forming a stack.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000479.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000025.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_016/sample_frames/left_camera/000025.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_016/sample_frames/right_camera/000025.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000102.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_016/sample_frames/left_camera/000102.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_016/sample_frames/right_camera/000102.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Align bowls the blue bowl in descending size, forming a stack.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Align bowls the blue bowl in descending size, forming a stack.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 19.4, "hop_absolute_value": 19.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_028-train-bf_000449-af_000464-plus-0000520", "task": "Use the left arm to grab red block, set it on the left, follow with the right arm for green block, and end with the left arm for blue block.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_012/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_012/sample_frames/head_camera/000466.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000449.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_028/sample_frames/left_camera/000449.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_028/sample_frames/right_camera/000449.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000464.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_028/sample_frames/left_camera/000464.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_028/sample_frames/right_camera/000464.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the left arm to grab red block, set it on the left, follow with the right arm for green block, and end with the left arm for blue block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the left arm to grab red block, set it on the left, follow with the right arm for green block, and end with the left arm for blue block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_three_episode_014-train-bf_000342-af_000411-plus-0000359", "task": "Relocate red block, green block, and blue block to the center and stack blue block on green block and green block on red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000480.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000342.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_014/sample_frames/left_camera/000342.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_014/sample_frames/right_camera/000342.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000411.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_014/sample_frames/left_camera/000411.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_014/sample_frames/right_camera/000411.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Relocate red block, green block, and blue block to the center and stack blue block on green block and green block on red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Relocate red block, green block, and blue block to the center and stack blue block on green block and green block on red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_019-train-bf_000025-af_000203-plus-0000324", "task": "Center red block and green block on the table and stack green block on top of red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_029/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_029/sample_frames/head_camera/000323.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000025.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_019/sample_frames/left_camera/000025.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_019/sample_frames/right_camera/000025.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000203.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_019/sample_frames/left_camera/000203.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_019/sample_frames/right_camera/000203.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Center red block and green block on the table and stack green block on top of red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Center red block and green block on the table and stack green block on top of red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 65.0, "hop_absolute_value": 65.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_mic_episode_002-train-bf_000065-af_000221-plus-0000258", "task": "Pick the white microphone handle black accents and transfer it to the other arm.", "image": [ "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_029/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_029/sample_frames/head_camera/000223.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000065.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000065.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000065.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000221.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000221.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000221.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the white microphone handle black accents and transfer it to the other arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the white microphone handle black accents and transfer it to the other arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_dual_shoes_episode_007-train-bf_000040-af_000196-plus-0000231", "task": "Put two the shoe with textured sole in the orange cardboard shoe-box with tips facing left.", "image": [ "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000226.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_007/sample_frames/head_camera/000040.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_007/sample_frames/left_camera/000040.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_007/sample_frames/right_camera/000040.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_007/sample_frames/head_camera/000196.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_007/sample_frames/left_camera/000196.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_007/sample_frames/right_camera/000196.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put two the shoe with textured sole in the orange cardboard shoe-box with tips facing left.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put two the shoe with textured sole in the orange cardboard shoe-box with tips facing left.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 78.6, "hop_absolute_value": 78.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_002-train-bf_000164-af_000259-plus-0000806", "task": "First, place red block, then green block beside it, and finally blue block next to green block.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000460.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000164.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000164.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000164.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000259.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000259.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000259.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`First, place red block, then green block beside it, and finally blue block next to green block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `First, place red block, then green block beside it, and finally blue block next to green block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_dual_shoes_episode_006-train-bf_000040-af_000196-plus-0000234", "task": "Place two the white shoe with flat sole inside the orange cardboard shoe-box, make tips face left.", "image": [ "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000224.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000040.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000040.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000040.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000196.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000196.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000196.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place two the white shoe with flat sole inside the orange cardboard shoe-box, make tips face left.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place two the white shoe with flat sole inside the orange cardboard shoe-box, make tips face left.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 78.6, "hop_absolute_value": 78.6, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_027-train-bf_000205-af_000245-plus-0000676", "task": "Move large block, medium block, and small block to the table center in size order.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000444.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000205.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000205.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000205.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000245.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000245.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000245.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move large block, medium block, and small block to the table center in size order.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move large block, medium block, and small block to the table center in size order.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_020-train-bf_000056-af_000631-plus-0000586", "task": "Grab the bottle with small ridged green cap, put it in the dustbin with lid and bag, follow by moving the clear bottle with blue sections and the hand-sized drinking bottle there too.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000660.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000056.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_020/sample_frames/left_camera/000056.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_020/sample_frames/right_camera/000056.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000631.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_020/sample_frames/left_camera/000631.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_020/sample_frames/right_camera/000631.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab the bottle with small ridged green cap, put it in the dustbin with lid and bag, follow by moving the clear bottle with blue sections and the hand-sized drinking bottle there too.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab the bottle with small ridged green cap, put it in the dustbin with lid and bag, follow by moving the clear bottle with blue sections and the hand-sized drinking bottle there too.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 97.4, "hop_absolute_value": 97.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_hanging_mug_episode_000-train-bf_000038-af_000337-plus-0000455", "task": "Pick the black coffee mug, turn it, put it in the center, and transfer it to the smooth dark gray rack.", "image": [ "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000337.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000038.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_000/sample_frames/left_camera/000038.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_000/sample_frames/right_camera/000038.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000337.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_000/sample_frames/left_camera/000337.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_000/sample_frames/right_camera/000337.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the black coffee mug, turn it, put it in the center, and transfer it to the smooth dark gray rack.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the black coffee mug, turn it, put it in the center, and transfer it to the smooth dark gray rack.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_shoe_episode_009-train-bf_000000-af_000150-plus-0000137", "task": "Transfer the medium red shoe from the table to the mat", "image": [ "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000187.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_009/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_009/sample_frames/right_camera/000000.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000150.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_009/sample_frames/left_camera/000150.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_009/sample_frames/right_camera/000150.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Transfer the medium red shoe from the table to the mat`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Transfer the medium red shoe from the table to the mat`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_hanging_mug_episode_016-train-bf_000000-af_000012-plus-0000472", "task": "Use your left arm to grab the dark green ceramic cup, rotate it, set it down, then hang on the angular rack with slanted support rod.", "image": [ "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000331.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_016/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_016/sample_frames/right_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000012.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_016/sample_frames/left_camera/000012.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_016/sample_frames/right_camera/000012.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use your left arm to grab the dark green ceramic cup, rotate it, set it down, then hang on the angular rack with slanted support rod.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use your left arm to grab the dark green ceramic cup, rotate it, set it down, then hang on the angular rack with slanted support rod.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 4.2, "hop_absolute_value": 4.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_hanging_mug_episode_012-train-bf_000228-af_000337-plus-0000434", "task": "Take the smooth ceramic mug, rotate it, set it in the middle, then hang it on the metal rack.", "image": [ "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000335.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_012/sample_frames/head_camera/000228.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_012/sample_frames/left_camera/000228.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_012/sample_frames/right_camera/000228.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_012/sample_frames/head_camera/000337.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_012/sample_frames/left_camera/000337.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_012/sample_frames/right_camera/000337.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the smooth ceramic mug, rotate it, set it in the middle, then hang it on the metal rack.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the smooth ceramic mug, rotate it, set it in the middle, then hang it on the metal rack.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_can_basket_episode_020-train-bf_000124-af_000217-plus-0000261", "task": "Place the red shiny can with green label in the basket with arched black grip and then lift the basket with arched black grip.", "image": [ "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000243.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000124.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_020/sample_frames/left_camera/000124.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_020/sample_frames/right_camera/000124.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000217.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_020/sample_frames/left_camera/000217.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_020/sample_frames/right_camera/000217.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the red shiny can with green label in the basket with arched black grip and then lift the basket with arched black grip.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the red shiny can with green label in the basket with arched black grip and then lift the basket with arched black grip.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 72.7, "hop_absolute_value": 72.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_object_cabinet_episode_011-train-bf_000063-af_000207-plus-0000234", "task": "Open the smooth-surface cabinet's drawer and transfer the small mouse with smooth surface into it.", "image": [ "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_001/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_001/sample_frames/head_camera/000268.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000063.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000063.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000063.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000207.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000207.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000207.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the smooth-surface cabinet's drawer and transfer the small mouse with smooth surface into it.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the smooth-surface cabinet's drawer and transfer the small mouse with smooth surface into it.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 53.3, "hop_absolute_value": 53.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_three_episode_023-train-bf_000167-af_000290-plus-0000902", "task": "Place the bottom the medium-sized bowl for holding food, then add the other two on top.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_023/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_023/sample_frames/head_camera/000467.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_023/sample_frames/head_camera/000167.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_023/sample_frames/left_camera/000167.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_023/sample_frames/right_camera/000167.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_023/sample_frames/head_camera/000290.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_023/sample_frames/left_camera/000290.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_023/sample_frames/right_camera/000290.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the bottom the medium-sized bowl for holding food, then add the other two on top.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the bottom the medium-sized bowl for holding food, then add the other two on top.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 45.0, "hop_absolute_value": 45.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_022-train-bf_000156-af_000172-plus-0000217", "task": "Relocate red block and green block to the center, then layer green block over red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_026/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_026/sample_frames/head_camera/000309.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000156.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_022/sample_frames/left_camera/000156.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_022/sample_frames/right_camera/000156.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000172.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_022/sample_frames/left_camera/000172.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_022/sample_frames/right_camera/000172.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Relocate red block and green block to the center, then layer green block over red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Relocate red block and green block to the center, then layer green block over red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 10.0, "hop_absolute_value": 10.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_002-train-bf_000109-af_000348-plus-0000452", "task": "Place large block, medium block, small block centrally and arrange largest to smallest.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000477.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000109.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_002/sample_frames/left_camera/000109.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_002/sample_frames/right_camera/000109.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000348.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_002/sample_frames/left_camera/000348.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_002/sample_frames/right_camera/000348.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place large block, medium block, small block centrally and arrange largest to smallest.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place large block, medium block, small block centrally and arrange largest to smallest.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 68.0, "hop_absolute_value": 68.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_dual_shoes_episode_011-train-bf_000000-af_000167-plus-0000202", "task": "Lift two the single red shoe, ensure left tips, drop in the shoe-box with Nike printed logo.", "image": [ "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000222.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000000.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000167.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000167.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000167.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift two the single red shoe, ensure left tips, drop in the shoe-box with Nike printed logo.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift two the single red shoe, ensure left tips, drop in the shoe-box with Nike printed logo.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 56.2, "hop_absolute_value": 56.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_three_episode_009-train-bf_000025-af_000263-plus-0000400", "task": "Move red block, green block, and blue block to the table's center and stack green block over red block, blue block over green block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000472.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000025.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_009/sample_frames/left_camera/000025.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_009/sample_frames/right_camera/000025.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000263.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_009/sample_frames/left_camera/000263.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_009/sample_frames/right_camera/000263.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move red block, green block, and blue block to the table's center and stack green block over red block, blue block over green block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move red block, green block, and blue block to the table's center and stack green block over red block, blue block over green block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 58.1, "hop_absolute_value": 58.1, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_005-train-bf_000045-af_000150-plus-0000296", "task": "Grab large block, medium block, and small block and align them by size at the center.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000462.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_005/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_005/sample_frames/right_camera/000045.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000150.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_005/sample_frames/left_camera/000150.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_005/sample_frames/right_camera/000150.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab large block, medium block, and small block and align them by size at the center.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab large block, medium block, and small block and align them by size at the center.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.9, "hop_absolute_value": 25.9, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_hanging_mug_episode_000-train-bf_000180-af_000242-plus-0000281", "task": "Use the left arm to grab the ceramic mug with cartoons, rotate it, place it in the middle, then use the right arm to hang it onto the metallic rack with smooth finish.", "image": [ "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000335.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000180.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_000/sample_frames/left_camera/000180.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_000/sample_frames/right_camera/000180.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000242.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_000/sample_frames/left_camera/000242.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_000/sample_frames/right_camera/000242.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the left arm to grab the ceramic mug with cartoons, rotate it, place it in the middle, then use the right arm to hang it onto the metallic rack with smooth finish.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the left arm to grab the ceramic mug with cartoons, rotate it, place it in the middle, then use the right arm to hang it onto the metallic rack with smooth finish.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_012-train-bf_000112-af_000235-plus-0000388", "task": "With arms, place the medium plastic bottle with tapered neck, the clear bottle with blue sections, and the red-labeled soda bottle into the trash bin step by step.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000657.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_012/sample_frames/head_camera/000112.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_012/sample_frames/left_camera/000112.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_012/sample_frames/right_camera/000112.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_012/sample_frames/head_camera/000235.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_012/sample_frames/left_camera/000235.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_012/sample_frames/right_camera/000235.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`With arms, place the medium plastic bottle with tapered neck, the clear bottle with blue sections, and the red-labeled soda bottle into the trash bin step by step.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `With arms, place the medium plastic bottle with tapered neck, the clear bottle with blue sections, and the red-labeled soda bottle into the trash bin step by step.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 48.5, "hop_absolute_value": 48.5, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_shake_bottle_episode_022-train-bf_000015-af_000060-plus-0000168", "task": "Pick the yellow bottle and perform shaking motion.", "image": [ "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000250.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000015.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_022/sample_frames/left_camera/000015.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_022/sample_frames/right_camera/000015.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000060.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_022/sample_frames/left_camera/000060.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_022/sample_frames/right_camera/000060.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the yellow bottle and perform shaking motion.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the yellow bottle and perform shaking motion.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 18.8, "hop_absolute_value": 18.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_002-train-bf_000086-af_000341-plus-0000467", "task": "Move the bottle made of smooth plastic, the bottle with ribbed design, and the black bottle with red label to the black garbage bin sequentially using the arms.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_011/sample_frames/head_camera/000669.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000086.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000086.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000086.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000341.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000341.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000341.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the bottle made of smooth plastic, the bottle with ribbed design, and the black bottle with red label to the black garbage bin sequentially using the arms.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the bottle made of smooth plastic, the bottle with ribbed design, and the black bottle with red label to the black garbage bin sequentially using the arms.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 71.0, "hop_absolute_value": 71.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_block_episode_029-train-bf_000057-af_000143-plus-0000244", "task": "Pick up the red block using the left arm, give it to the right arm, and place it on the blue pad.", "image": [ "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_029/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_029/sample_frames/head_camera/000299.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_029/sample_frames/head_camera/000057.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_029/sample_frames/left_camera/000057.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_029/sample_frames/right_camera/000057.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_029/sample_frames/head_camera/000143.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_029/sample_frames/left_camera/000143.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_029/sample_frames/right_camera/000143.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the red block using the left arm, give it to the right arm, and place it on the blue pad.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the red block using the left arm, give it to the right arm, and place it on the blue pad.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 35.3, "hop_absolute_value": 35.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_017-train-bf_000165-af_000390-plus-0000428", "task": "Handle the microwave with knobs and open it with the left arm.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000487.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000165.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_017/sample_frames/left_camera/000165.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_017/sample_frames/right_camera/000165.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000390.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_017/sample_frames/left_camera/000390.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_017/sample_frames/right_camera/000390.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Handle the microwave with knobs and open it with the left arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Handle the microwave with knobs and open it with the left arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 71.4, "hop_absolute_value": 71.4, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_two_episode_004-train-bf_000038-af_000115-plus-0000245", "task": "Grab the blue bowl, then stack the rounded bowl with light blue color on top.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000315.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000038.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_004/sample_frames/left_camera/000038.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_004/sample_frames/right_camera/000038.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000115.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_004/sample_frames/left_camera/000115.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_004/sample_frames/right_camera/000115.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab the blue bowl, then stack the rounded bowl with light blue color on top.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab the blue bowl, then stack the rounded bowl with light blue color on top.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 31.6, "hop_absolute_value": 31.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_001-train-bf_000359-af_000633-plus-0000523", "task": "Use arms to move the medium plastic bottle with tapered neck, the medium-sized water bottle, and the black soda bottle with white top into the plastic dustbin.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000666.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_001/sample_frames/head_camera/000359.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_001/sample_frames/left_camera/000359.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_001/sample_frames/right_camera/000359.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_001/sample_frames/head_camera/000633.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_001/sample_frames/left_camera/000633.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_001/sample_frames/right_camera/000633.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use arms to move the medium plastic bottle with tapered neck, the medium-sized water bottle, and the black soda bottle with white top into the plastic dustbin.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use arms to move the medium plastic bottle with tapered neck, the medium-sized water bottle, and the black soda bottle with white top into the plastic dustbin.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 90.9, "hop_absolute_value": 90.9, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_move_stapler_pad_episode_005-train-bf_000090-af_000105-plus-0000060", "task": "Drop the small red stapler onto the Blue mat.", "image": [ "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000152.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000090.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_clean_50/episode_005/sample_frames/left_camera/000090.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_clean_50/episode_005/sample_frames/right_camera/000090.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000105.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_clean_50/episode_005/sample_frames/left_camera/000105.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_clean_50/episode_005/sample_frames/right_camera/000105.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Drop the small red stapler onto the Blue mat.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Drop the small red stapler onto the Blue mat.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_000-train-bf_000150-af_000177-plus-0000605", "task": "Use the right arm, the left arm, and the right arm to arrange large block, medium block, and small block centrally by size.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000474.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000150.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_000/sample_frames/left_camera/000150.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_000/sample_frames/right_camera/000150.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000177.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_000/sample_frames/left_camera/000177.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_000/sample_frames/right_camera/000177.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the right arm, the left arm, and the right arm to arrange large block, medium block, and small block centrally by size.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the right arm, the left arm, and the right arm to arrange large block, medium block, and small block centrally by size.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.1, "hop_absolute_value": 9.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_015-train-bf_000082-af_000300-plus-0000392", "task": "Using the right arm, the left arm, and the right arm, align red block, green block, and blue block in a row from left to right.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000465.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000082.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_015/sample_frames/left_camera/000082.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_015/sample_frames/right_camera/000082.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000300.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_015/sample_frames/left_camera/000300.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_015/sample_frames/right_camera/000300.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Using the right arm, the left arm, and the right arm, align red block, green block, and blue block in a row from left to right.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Using the right arm, the left arm, and the right arm, align red block, green block, and blue block in a row from left to right.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 59.3, "hop_absolute_value": 59.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_008-train-bf_000189-af_000206-plus-0000222", "task": "Use the right arm to move red block, then the left arm to stack green block on red block at the center.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_017/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_017/sample_frames/head_camera/000323.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_008/sample_frames/head_camera/000189.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_008/sample_frames/left_camera/000189.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_008/sample_frames/right_camera/000189.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_008/sample_frames/head_camera/000206.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_008/sample_frames/left_camera/000206.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_008/sample_frames/right_camera/000206.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the right arm to move red block, then the left arm to stack green block on red block at the center.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the right arm to move red block, then the left arm to stack green block on red block at the center.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 12.5, "hop_absolute_value": 12.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_press_stapler_episode_008-train-bf_000000-af_000030-plus-0000041", "task": "Firmly push down on the manual blue stapler.", "image": [ "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000121.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_randomized_500/episode_008/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_randomized_500/episode_008/sample_frames/right_camera/000000.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000030.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_randomized_500/episode_008/sample_frames/left_camera/000030.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_randomized_500/episode_008/sample_frames/right_camera/000030.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Firmly push down on the manual blue stapler.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Firmly push down on the manual blue stapler.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_shoe_episode_013-train-bf_000000-af_000120-plus-0000124", "task": "Pick the white shoe with three stripes up from the table and put it on the mat", "image": [ "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000174.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_013/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_013/sample_frames/right_camera/000000.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000120.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_013/sample_frames/left_camera/000120.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_013/sample_frames/right_camera/000120.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the white shoe with three stripes up from the table and put it on the mat`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the white shoe with three stripes up from the table and put it on the mat`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_023-train-bf_000089-af_000155-plus-0000490", "task": "Move red block and green block to the center; stack green block on top of red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000316.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_023/sample_frames/head_camera/000089.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_023/sample_frames/left_camera/000089.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_023/sample_frames/right_camera/000089.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_023/sample_frames/head_camera/000155.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_023/sample_frames/left_camera/000155.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_023/sample_frames/right_camera/000155.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move red block and green block to the center; stack green block on top of red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move red block and green block to the center; stack green block on top of red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_shake_bottle_episode_001-train-bf_000000-af_000045-plus-0000302", "task": "Shake the the cylindrical bottle with narrow neck properly after lifting.", "image": [ "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000254.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_001/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_001/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_001/sample_frames/right_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_001/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_001/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_001/sample_frames/right_camera/000045.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Shake the the cylindrical bottle with narrow neck properly after lifting.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Shake the the cylindrical bottle with narrow neck properly after lifting.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 17.6, "hop_absolute_value": 17.6, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_burger_fries_episode_029-train-bf_000033-af_000170-plus-0000237", "task": "Pick up the round, soft bread hamburg, then the Golden fries crispy red package container, and place them onto the smooth rectangular tray.", "image": [ "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000240.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_029/sample_frames/head_camera/000033.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_029/sample_frames/left_camera/000033.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_029/sample_frames/right_camera/000033.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_029/sample_frames/head_camera/000170.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_029/sample_frames/left_camera/000170.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_029/sample_frames/right_camera/000170.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the round, soft bread hamburg, then the Golden fries crispy red package container, and place them onto the smooth rectangular tray.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the round, soft bread hamburg, then the Golden fries crispy red package container, and place them onto the smooth rectangular tray.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 56.2, "hop_absolute_value": 56.2, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_mic_episode_002-train-bf_000049-af_000114-plus-0000164", "task": "Pick the white microphone handle black accents and transfer it to the other arm.", "image": [ "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000222.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000049.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000049.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000049.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000114.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000114.png", "images/robotwin2_agilex_part1/handover_mic_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000114.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the white microphone handle black accents and transfer it to the other arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the white microphone handle black accents and transfer it to the other arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 30.8, "hop_absolute_value": 30.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_028-train-bf_000205-af_000286-plus-0000797", "task": "Put large block, medium block, and small block in the center by size.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000462.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000205.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/left_camera/000205.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/right_camera/000205.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000286.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/left_camera/000286.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/right_camera/000286.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put large block, medium block, and small block in the center by size.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put large block, medium block, and small block in the center by size.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_019-train-bf_000075-af_000165-plus-0000811", "task": "Access the inside by opening the dark gray kitchen microwave.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000432.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000075.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_019/sample_frames/left_camera/000075.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_019/sample_frames/right_camera/000075.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000165.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_019/sample_frames/left_camera/000165.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_019/sample_frames/right_camera/000165.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Access the inside by opening the dark gray kitchen microwave.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Access the inside by opening the dark gray kitchen microwave.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 16.7, "hop_absolute_value": 16.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_021-train-bf_000000-af_000103-plus-0000330", "task": "Use the arms to move the yellow bottle body with tapered neck into the black bin with white liner, then repeat for the cylindrical bottle and the black soda bottle with white top.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000661.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_021/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_021/sample_frames/right_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000103.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_021/sample_frames/left_camera/000103.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_021/sample_frames/right_camera/000103.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the arms to move the yellow bottle body with tapered neck into the black bin with white liner, then repeat for the cylindrical bottle and the black soda bottle with white top.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the arms to move the yellow bottle body with tapered neck into the black bin with white liner, then repeat for the cylindrical bottle and the black soda bottle with white top.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 24.4, "hop_absolute_value": 24.4, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_014-train-bf_000014-af_000123-plus-0000752", "task": "Bring large block, medium block, small block to the table's center and position them left to right.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000463.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000014.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_014/sample_frames/left_camera/000014.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_014/sample_frames/right_camera/000014.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000123.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_014/sample_frames/left_camera/000123.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_014/sample_frames/right_camera/000123.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Bring large block, medium block, small block to the table's center and position them left to right.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Bring large block, medium block, small block to the table's center and position them left to right.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_019-train-bf_000102-af_000203-plus-0000283", "task": "Use the left arm to center red block, then stack green block with the right arm.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000319.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000102.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000102.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000102.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000203.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000203.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000203.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the left arm to center red block, then stack green block with the right arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the left arm to center red block, then stack green block with the right arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_014-train-bf_000000-af_000082-plus-0000677", "task": "Put large block, medium block, and small block in the center by size.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000456.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_014/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_014/sample_frames/right_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000082.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_014/sample_frames/left_camera/000082.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_014/sample_frames/right_camera/000082.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put large block, medium block, and small block in the center by size.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put large block, medium block, and small block in the center by size.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 18.2, "hop_absolute_value": 18.2, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_019-train-bf_000300-af_000408-plus-0000431", "task": "Pick large block, medium block, and small block sequentially and move them to the center.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000449.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000300.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000300.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000300.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000408.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000408.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000408.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick large block, medium block, and small block sequentially and move them to the center.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick large block, medium block, and small block sequentially and move them to the center.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 72.7, "hop_absolute_value": 72.7, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_013-train-bf_000109-af_000191-plus-0000761", "task": "Grab large block, medium block, and small block, set them centrally, ordered by size.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000473.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000109.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_013/sample_frames/left_camera/000109.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_013/sample_frames/right_camera/000109.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000191.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_013/sample_frames/left_camera/000191.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_clean_50/episode_013/sample_frames/right_camera/000191.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab large block, medium block, and small block, set them centrally, ordered by size.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab large block, medium block, and small block, set them centrally, ordered by size.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 24.0, "hop_absolute_value": 24.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_move_stapler_pad_episode_013-train-bf_000045-af_000105-plus-0000092", "task": "Place the blue stapler with silver staple tray on the Red mat using the left arm", "image": [ "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000157.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_randomized_500/episode_013/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_randomized_500/episode_013/sample_frames/right_camera/000045.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000105.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_randomized_500/episode_013/sample_frames/left_camera/000105.png", "images/robotwin2_agilex_part1/move_stapler_pad_aloha-agilex_randomized_500/episode_013/sample_frames/right_camera/000105.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the blue stapler with silver staple tray on the Red mat using the left arm`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the blue stapler with silver staple tray on the Red mat using the left arm`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_010-train-bf_000120-af_000150-plus-0000572", "task": "Utilize the left arm to pull the gray microwave open.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000428.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000120.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_010/sample_frames/left_camera/000120.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_010/sample_frames/right_camera/000120.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000150.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_010/sample_frames/left_camera/000150.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_010/sample_frames/right_camera/000150.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Utilize the left arm to pull the gray microwave open.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Utilize the left arm to pull the gray microwave open.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 9.5, "hop_absolute_value": 9.5, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_cans_plasticbox_episode_028-train-bf_000150-af_000232-plus-0000285", "task": "Place the metallic soda can and the green red metallic soda can into the durable plasticbox with side handle with dual arms", "image": [ "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000286.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000150.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_028/sample_frames/left_camera/000150.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_028/sample_frames/right_camera/000150.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000232.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_028/sample_frames/left_camera/000232.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_028/sample_frames/right_camera/000232.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the metallic soda can and the green red metallic soda can into the durable plasticbox with side handle with dual arms`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the metallic soda can and the green red metallic soda can into the durable plasticbox with side handle with dual arms`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_011-train-bf_000038-af_000312-plus-0000411", "task": "Shift red block and green block to the middle, then stack green block over red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000312.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000038.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000038.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000038.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000312.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000312.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000312.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Shift red block and green block to the middle, then stack green block over red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Shift red block and green block to the middle, then stack green block over red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_two_episode_008-train-bf_000089-af_000215-plus-0000315", "task": "Position the glazed bowl with curved sides and stack the blue bowl securely onto it.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000305.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000089.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_008/sample_frames/left_camera/000089.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_008/sample_frames/right_camera/000089.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000215.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_008/sample_frames/left_camera/000215.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_008/sample_frames/right_camera/000215.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Position the glazed bowl with curved sides and stack the blue bowl securely onto it.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Position the glazed bowl with curved sides and stack the blue bowl securely onto it.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_dump_bin_bigbin_episode_014-train-bf_000060-af_000165-plus-0000264", "task": "Pour balls from the light gray tabletrashbin into the bin.", "image": [ "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000174.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000060.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_014/sample_frames/left_camera/000060.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_014/sample_frames/right_camera/000060.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000165.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_014/sample_frames/left_camera/000165.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_014/sample_frames/right_camera/000165.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pour balls from the light gray tabletrashbin into the bin.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pour balls from the light gray tabletrashbin into the bin.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 38.9, "hop_absolute_value": 38.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_two_episode_023-train-bf_000127-af_000201-plus-0000261", "task": "Take the ceramic bowl, place it, and stack the blue bowl with slightly rounded base on top.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_001/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_001/sample_frames/head_camera/000331.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_023/sample_frames/head_camera/000127.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_023/sample_frames/left_camera/000127.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_023/sample_frames/right_camera/000127.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_023/sample_frames/head_camera/000201.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_023/sample_frames/left_camera/000201.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_023/sample_frames/right_camera/000201.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the ceramic bowl, place it, and stack the blue bowl with slightly rounded base on top.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the ceramic bowl, place it, and stack the blue bowl with slightly rounded base on top.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 41.7, "hop_absolute_value": 41.7, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_cans_plasticbox_episode_014-train-bf_000045-af_000177-plus-0000273", "task": "Use both arms, position the colorful can with lime and orange design and the red can with gold illustrations within the rectangular plasticbox with rounded edges.", "image": [ "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000285.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_014/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_014/sample_frames/right_camera/000045.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000177.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_014/sample_frames/left_camera/000177.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_clean_50/episode_014/sample_frames/right_camera/000177.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use both arms, position the colorful can with lime and orange design and the red can with gold illustrations within the rectangular plasticbox with rounded edges.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use both arms, position the colorful can with lime and orange design and the red can with gold illustrations within the rectangular plasticbox with rounded edges.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 52.9, "hop_absolute_value": 52.9, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_can_basket_episode_018-train-bf_000188-af_000243-plus-0000304", "task": "Pick up the red beverage can with curved top, set it inside the basket made of woven wood, then lift the basket made of woven wood.", "image": [ "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_011/sample_frames/head_camera/000267.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000188.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_018/sample_frames/left_camera/000188.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_018/sample_frames/right_camera/000188.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000243.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_018/sample_frames/left_camera/000243.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_randomized_500/episode_018/sample_frames/right_camera/000243.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the red beverage can with curved top, set it inside the basket made of woven wood, then lift the basket made of woven wood.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the red beverage can with curved top, set it inside the basket made of woven wood, then lift the basket made of woven wood.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_object_cabinet_episode_005-train-bf_000248-af_000256-plus-0000216", "task": "Open the drawer of the wooden cabinet before placing the phone with digital display inside it.", "image": [ "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000266.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000248.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_005/sample_frames/left_camera/000248.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_005/sample_frames/right_camera/000248.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000256.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_005/sample_frames/left_camera/000256.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_005/sample_frames/right_camera/000256.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the drawer of the wooden cabinet before placing the phone with digital display inside it.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the drawer of the wooden cabinet before placing the phone with digital display inside it.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "progress", "combined_category": "progress_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_005-train-bf_000210-af_000225-plus-0000554", "task": "Use the left arm to pull the microwave with knobs open.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000431.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000210.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_005/sample_frames/left_camera/000210.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_005/sample_frames/right_camera/000210.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000225.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_005/sample_frames/left_camera/000225.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_005/sample_frames/right_camera/000225.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the left arm to pull the microwave with knobs open.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the left arm to pull the microwave with knobs open.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 6.7, "hop_absolute_value": 6.7, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_003-train-bf_000028-af_000103-plus-0000859", "task": "Use the arms to transfer the smooth yellow bottle with ridged cap, the bottle with narrow neck, and the bottle with curved body shape into the dustbin with white inner bag step by step.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000660.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000028.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_003/sample_frames/left_camera/000028.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_003/sample_frames/right_camera/000028.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000103.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_003/sample_frames/left_camera/000103.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_003/sample_frames/right_camera/000103.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the arms to transfer the smooth yellow bottle with ridged cap, the bottle with narrow neck, and the bottle with curved body shape into the dustbin with white inner bag step by step.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the arms to transfer the smooth yellow bottle with ridged cap, the bottle with narrow neck, and the bottle with curved body shape into the dustbin with white inner bag step by step.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 19.0, "hop_absolute_value": 19.0, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_016-train-bf_000117-af_000128-plus-0000284", "task": "Place the yellow bottle with smooth texture, the bottle with turquoise cap, and the bottle with curved body shape into the medium-sized black plastic dustbin one after another.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000524.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000117.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_016/sample_frames/left_camera/000117.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_016/sample_frames/right_camera/000117.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000128.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_016/sample_frames/left_camera/000128.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_016/sample_frames/right_camera/000128.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the yellow bottle with smooth texture, the bottle with turquoise cap, and the bottle with curved body shape into the medium-sized black plastic dustbin one after another.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the yellow bottle with smooth texture, the bottle with turquoise cap, and the bottle with curved body shape into the medium-sized black plastic dustbin one after another.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 3.8, "hop_absolute_value": 3.8, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_bread_basket_episode_008-train-bf_000300-af_000313-plus-0000371", "task": "Lift the bread loaf with uneven edges and the small golden baked bread at once with the left arm, then set them in the smooth-textured breadbasket.", "image": [ "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000228.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000300.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_randomized_500/episode_008/sample_frames/left_camera/000300.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_randomized_500/episode_008/sample_frames/right_camera/000300.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000313.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_randomized_500/episode_008/sample_frames/left_camera/000313.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_randomized_500/episode_008/sample_frames/right_camera/000313.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift the bread loaf with uneven edges and the small golden baked bread at once with the left arm, then set them in the smooth-textured breadbasket.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift the bread loaf with uneven edges and the small golden baked bread at once with the left arm, then set them in the smooth-textured breadbasket.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "progress", "combined_category": "progress_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_023-train-bf_000240-af_000255-plus-0000564", "task": "Open the microwave by pulling on its handle.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000421.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_023/sample_frames/head_camera/000240.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_023/sample_frames/left_camera/000240.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_023/sample_frames/right_camera/000240.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_023/sample_frames/head_camera/000255.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_023/sample_frames/left_camera/000255.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_023/sample_frames/right_camera/000255.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the microwave by pulling on its handle.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the microwave by pulling on its handle.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "+1" } ], "hop_value": 7.1, "hop_absolute_value": 7.1, "hop_category": { "absolute_category": "small", "direction": "progress", "combined_category": "progress_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_block_episode_020-train-bf_000237-af_000143-minus-0000147", "task": "Grab red block, pass to the right arm, then drop on blue pad.", "image": [ "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000276.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000237.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_020/sample_frames/left_camera/000237.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_020/sample_frames/right_camera/000237.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000143.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_020/sample_frames/left_camera/000143.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_020/sample_frames/right_camera/000143.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab red block, pass to the right arm, then drop on blue pad.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab red block, pass to the right arm, then drop on blue pad.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -41.2, "hop_absolute_value": 41.2, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_hanging_mug_episode_017-train-bf_000150-af_000025-minus-0000060", "task": "Take the hand-sized mug with blue artwork, rotate it, place it in the center, then hang it on the rack with angular body.", "image": [ "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_017/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_017/sample_frames/head_camera/000333.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_017/sample_frames/head_camera/000150.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_017/sample_frames/left_camera/000150.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_017/sample_frames/right_camera/000150.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_017/sample_frames/head_camera/000025.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_017/sample_frames/left_camera/000025.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_017/sample_frames/right_camera/000025.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the hand-sized mug with blue artwork, rotate it, place it in the center, then hang it on the rack with angular body.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the hand-sized mug with blue artwork, rotate it, place it in the center, then hang it on the rack with angular body.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_shake_bottle_horizontally_episode_018-train-bf_000279-af_000255-minus-0000169", "task": "Pick up the cylindrical bottle with narrow neck using the left arm, shake it horizontally.", "image": [ "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000282.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000279.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_018/sample_frames/left_camera/000279.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_018/sample_frames/right_camera/000279.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000255.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_018/sample_frames/left_camera/000255.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_018/sample_frames/right_camera/000255.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the cylindrical bottle with narrow neck using the left arm, shake it horizontally.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the cylindrical bottle with narrow neck using the left arm, shake it horizontally.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -10.5, "hop_absolute_value": 10.5, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_phone_stand_episode_012-train-bf_000105-af_000075-minus-0000034", "task": "Move the sleek black phone onto the phonestand with smooth texture.", "image": [ "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_011/sample_frames/head_camera/000127.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_012/sample_frames/head_camera/000105.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_012/sample_frames/left_camera/000105.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_012/sample_frames/right_camera/000105.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_012/sample_frames/head_camera/000075.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_012/sample_frames/left_camera/000075.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_012/sample_frames/right_camera/000075.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the sleek black phone onto the phonestand with smooth texture.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the sleek black phone onto the phonestand with smooth texture.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -28.6, "hop_absolute_value": 28.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_three_episode_004-train-bf_000340-af_000195-minus-0000217", "task": "Place red block, green block, and blue block at the center, using the left arm, the left arm, and the right arm to stack blue block on green block and green block on red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_025/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_025/sample_frames/head_camera/000454.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_004/sample_frames/head_camera/000340.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_004/sample_frames/left_camera/000340.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_004/sample_frames/right_camera/000340.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_004/sample_frames/head_camera/000195.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_004/sample_frames/left_camera/000195.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_004/sample_frames/right_camera/000195.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place red block, green block, and blue block at the center, using the left arm, the left arm, and the right arm to stack blue block on green block and green block on red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place red block, green block, and blue block at the center, using the left arm, the left arm, and the right arm to stack blue block on green block and green block on red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_shoe_episode_018-train-bf_000150-af_000075-minus-0000048", "task": "Take the the rubber sole shoe off the table and put it on the mat", "image": [ "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000176.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000150.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_018/sample_frames/left_camera/000150.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_018/sample_frames/right_camera/000150.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000075.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_018/sample_frames/left_camera/000075.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_randomized_500/episode_018/sample_frames/right_camera/000075.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the the rubber sole shoe off the table and put it on the mat`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the the rubber sole shoe off the table and put it on the mat`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_bread_skillet_episode_024-train-bf_000153-af_000128-minus-0000057", "task": "Use the left arm to grab the bread with rough texture and set it in the flat-bottomed skillet with long handle.", "image": [ "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000160.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000153.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_024/sample_frames/left_camera/000153.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_024/sample_frames/right_camera/000153.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000128.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_024/sample_frames/left_camera/000128.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_024/sample_frames/right_camera/000128.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the left arm to grab the bread with rough texture and set it in the flat-bottomed skillet with long handle.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the left arm to grab the bread with rough texture and set it in the flat-bottomed skillet with long handle.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -30.0, "hop_absolute_value": 30.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_mouse_pad_episode_000-train-bf_000135-af_000000-minus-0000009", "task": "Using the left arm, grab the the compact matte gray computer mouse and set it on the Yellow mat.", "image": [ "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_clean_50/episode_005/sample_frames/head_camera/000147.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000135.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_clean_50/episode_000/sample_frames/left_camera/000135.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_clean_50/episode_000/sample_frames/right_camera/000135.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_clean_50/episode_000/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_clean_50/episode_000/sample_frames/right_camera/000000.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Using the left arm, grab the the compact matte gray computer mouse and set it on the Yellow mat.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Using the left arm, grab the the compact matte gray computer mouse and set it on the Yellow mat.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_028-train-bf_000455-af_000232-minus-0000888", "task": "Put large block, medium block, and small block in the center by size.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000470.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000455.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/left_camera/000455.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/right_camera/000455.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000232.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/left_camera/000232.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_028/sample_frames/right_camera/000232.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Put large block, medium block, and small block in the center by size.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Put large block, medium block, and small block in the center by size.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -46.9, "hop_absolute_value": 46.9, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_019-train-bf_000037-af_000009-minus-0000117", "task": "Use the arms to transfer the yellow bottle with smooth texture, the smooth transparent plastic bottle, and the black and red bottle into the plastic dustbin step by step.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000649.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000037.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000037.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000037.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000009.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000009.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000009.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the arms to transfer the yellow bottle with smooth texture, the smooth transparent plastic bottle, and the black and red bottle into the plastic dustbin step by step.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the arms to transfer the yellow bottle with smooth texture, the smooth transparent plastic bottle, and the black and red bottle into the plastic dustbin step by step.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_two_episode_019-train-bf_000218-af_000102-minus-0000142", "task": "Position the curved bowl light blue shade carefully on the rounded bowl with light blue color.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_003/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_003/sample_frames/head_camera/000317.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000218.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000218.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000218.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000102.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000102.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000102.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Position the curved bowl light blue shade carefully on the rounded bowl with light blue color.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Position the curved bowl light blue shade carefully on the rounded bowl with light blue color.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_beat_block_hammer_episode_019-train-bf_000090-af_000030-minus-0000016", "task": "Pick up the plastic handle metal hammer using the left arm, hammer the block.", "image": [ "images/robotwin2_agilex_part1/beat_block_hammer_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/beat_block_hammer_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000111.png", "images/robotwin2_agilex_part1/beat_block_hammer_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000090.png", "images/robotwin2_agilex_part1/beat_block_hammer_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000090.png", "images/robotwin2_agilex_part1/beat_block_hammer_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000090.png", "images/robotwin2_agilex_part1/beat_block_hammer_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000030.png", "images/robotwin2_agilex_part1/beat_block_hammer_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000030.png", "images/robotwin2_agilex_part1/beat_block_hammer_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000030.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the plastic handle metal hammer using the left arm, hammer the block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the plastic handle metal hammer using the left arm, hammer the block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_three_episode_008-train-bf_000375-af_000358-minus-0000274", "task": "Take each the glazed bowl with curved sides and place them one by one on top.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000463.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_008/sample_frames/head_camera/000375.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_008/sample_frames/left_camera/000375.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_008/sample_frames/right_camera/000375.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_008/sample_frames/head_camera/000358.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_008/sample_frames/left_camera/000358.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_008/sample_frames/right_camera/000358.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take each the glazed bowl with curved sides and place them one by one on top.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take each the glazed bowl with curved sides and place them one by one on top.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -3.7, "hop_absolute_value": 3.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_bread_basket_episode_015-train-bf_000120-af_000060-minus-0000067", "task": "Use two arms and set the bread loaf with uneven edges, the bread with warm golden color in the breadbasket shaped like a bowl.", "image": [ "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_012/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_012/sample_frames/head_camera/000226.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000120.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_015/sample_frames/left_camera/000120.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_015/sample_frames/right_camera/000120.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000060.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_015/sample_frames/left_camera/000060.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_015/sample_frames/right_camera/000060.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use two arms and set the bread loaf with uneven edges, the bread with warm golden color in the breadbasket shaped like a bowl.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use two arms and set the bread loaf with uneven edges, the bread with warm golden color in the breadbasket shaped like a bowl.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_025-train-bf_000385-af_000314-minus-0000652", "task": "Grab large block, medium block, small block, and arrange them in size order at the center.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000459.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_025/sample_frames/head_camera/000385.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_025/sample_frames/left_camera/000385.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_025/sample_frames/right_camera/000385.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_025/sample_frames/head_camera/000314.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_025/sample_frames/left_camera/000314.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_025/sample_frames/right_camera/000314.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab large block, medium block, small block, and arrange them in size order at the center.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab large block, medium block, small block, and arrange them in size order at the center.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -17.9, "hop_absolute_value": 17.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_021-train-bf_000064-af_000051-minus-0000451", "task": "Bring red block to the center, then stack green block over it.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000316.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000064.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_021/sample_frames/left_camera/000064.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_021/sample_frames/right_camera/000064.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000051.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_021/sample_frames/left_camera/000051.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_021/sample_frames/right_camera/000051.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Bring red block to the center, then stack green block over it.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Bring red block to the center, then stack green block over it.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_027-train-bf_000238-af_000089-minus-0000115", "task": "Place red block and green block centrally, then stack green block on red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000319.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000238.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000238.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000238.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000089.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000089.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000089.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place red block and green block centrally, then stack green block on red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place red block and green block centrally, then stack green block on red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -58.8, "hop_absolute_value": 58.8, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_burger_fries_episode_014-train-bf_000150-af_000050-minus-0000069", "task": "Move the rectangular hamburg box and the red box with golden sticks to the medium bright orange tray with both arms", "image": [ "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000242.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000150.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_014/sample_frames/left_camera/000150.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_014/sample_frames/right_camera/000150.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000050.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_014/sample_frames/left_camera/000050.png", "images/robotwin2_agilex_part1/place_burger_fries_aloha-agilex_randomized_500/episode_014/sample_frames/right_camera/000050.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move the rectangular hamburg box and the red box with golden sticks to the medium bright orange tray with both arms`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move the rectangular hamburg box and the red box with golden sticks to the medium bright orange tray with both arms`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_three_episode_013-train-bf_000263-af_000115-minus-0000170", "task": "Start with the first the rounded bowl with light blue color, adding the next two in sequence.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000462.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000263.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_013/sample_frames/left_camera/000263.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_013/sample_frames/right_camera/000263.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000115.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_013/sample_frames/left_camera/000115.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_013/sample_frames/right_camera/000115.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Start with the first the rounded bowl with light blue color, adding the next two in sequence.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Start with the first the rounded bowl with light blue color, adding the next two in sequence.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -55.0, "hop_absolute_value": 55.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_019-train-bf_000464-af_000225-minus-0000197", "task": "Pull the handle to open the microwave with textured top panel.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_001/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_001/sample_frames/head_camera/000450.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000464.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000464.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000464.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000225.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000225.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000225.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the handle to open the microwave with textured top panel.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the handle to open the microwave with textured top panel.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -51.6, "hop_absolute_value": 51.6, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_006-train-bf_000374-af_000259-minus-0000231", "task": "Using the right arm, the left arm, and the right arm, arrange red block, green block, and blue block sequentially from left to right in a single row.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000451.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000374.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000374.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000374.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000259.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000259.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000259.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Using the right arm, the left arm, and the right arm, arrange red block, green block, and blue block sequentially from left to right in a single row.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Using the right arm, the left arm, and the right arm, arrange red block, green block, and blue block sequentially from left to right in a single row.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -29.6, "hop_absolute_value": 29.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_028-train-bf_000171-af_000064-minus-0000102", "task": "Move red block to the middle and put green block on top of it.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_025/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_025/sample_frames/head_camera/000314.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000171.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_028/sample_frames/left_camera/000171.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_028/sample_frames/right_camera/000171.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000064.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_028/sample_frames/left_camera/000064.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_randomized_500/episode_028/sample_frames/right_camera/000064.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move red block to the middle and put green block on top of it.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move red block to the middle and put green block on top of it.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -61.5, "hop_absolute_value": 61.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_block_episode_014-train-bf_000266-af_000171-minus-0000148", "task": "Lift, switch arms, and place the red block.", "image": [ "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000288.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000266.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_014/sample_frames/left_camera/000266.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_014/sample_frames/right_camera/000266.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000171.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_014/sample_frames/left_camera/000171.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_014/sample_frames/right_camera/000171.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift, switch arms, and place the red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift, switch arms, and place the red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_015-train-bf_000447-af_000041-minus-0000061", "task": "Move large block, medium block, and small block to the center and align.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_015/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_015/sample_frames/head_camera/000462.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_015/sample_frames/head_camera/000447.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_015/sample_frames/left_camera/000447.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_015/sample_frames/right_camera/000447.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_015/sample_frames/head_camera/000041.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_015/sample_frames/left_camera/000041.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_015/sample_frames/right_camera/000041.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move large block, medium block, and small block to the center and align.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move large block, medium block, and small block to the center and align.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -90.6, "hop_absolute_value": 90.6, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_001-train-bf_000495-af_000468-minus-0000610", "task": "Place the yellow bottle, the medium-sized water bottle, and the rounded bottle with a white neck into the black garbage bin one after another.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000660.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_001/sample_frames/head_camera/000495.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_001/sample_frames/left_camera/000495.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_001/sample_frames/right_camera/000495.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_001/sample_frames/head_camera/000468.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_001/sample_frames/left_camera/000468.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_001/sample_frames/right_camera/000468.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the yellow bottle, the medium-sized water bottle, and the rounded bottle with a white neck into the black garbage bin one after another.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the yellow bottle, the medium-sized water bottle, and the rounded bottle with a white neck into the black garbage bin one after another.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -2.6, "hop_absolute_value": 2.6, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_laptop_episode_006-train-bf_000150-af_000015-minus-0000015", "task": "Fully open the the compact silver laptop.", "image": [ "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_012/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_012/sample_frames/head_camera/000227.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000150.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000150.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000150.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000015.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000015.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000015.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Fully open the the compact silver laptop.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Fully open the the compact silver laptop.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -90.0, "hop_absolute_value": 90.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_000-train-bf_000245-af_000109-minus-0000923", "task": "Take large block with the right arm, medium block with the left arm, and small block with the right arm to the middle.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000473.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000245.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_000/sample_frames/left_camera/000245.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_000/sample_frames/right_camera/000245.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000109.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_000/sample_frames/left_camera/000109.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_000/sample_frames/right_camera/000109.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take large block with the right arm, medium block with the left arm, and small block with the right arm to the middle.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take large block with the right arm, medium block with the left arm, and small block with the right arm to the middle.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -55.6, "hop_absolute_value": 55.6, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_two_episode_012-train-bf_000155-af_000076-minus-0000135", "task": "Pick the blue bowl with slightly rounded base, then stack the blue bowl on it.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000301.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_012/sample_frames/head_camera/000155.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_012/sample_frames/left_camera/000155.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_012/sample_frames/right_camera/000155.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_012/sample_frames/head_camera/000076.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_012/sample_frames/left_camera/000076.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_012/sample_frames/right_camera/000076.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the blue bowl with slightly rounded base, then stack the blue bowl on it.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the blue bowl with slightly rounded base, then stack the blue bowl on it.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_hanging_mug_episode_010-train-bf_000050-af_000012-minus-0000081", "task": "Take the drinking mug shiny black exterior, rotate it, place it in the center, then hang it on the rack with two rods sticking out.", "image": [ "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000335.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000050.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/left_camera/000050.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/right_camera/000050.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000012.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/left_camera/000012.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/right_camera/000012.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the drinking mug shiny black exterior, rotate it, place it in the center, then hang it on the rack with two rods sticking out.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the drinking mug shiny black exterior, rotate it, place it in the center, then hang it on the rack with two rods sticking out.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_shoe_episode_010-train-bf_000045-af_000000-minus-0000000", "task": "Lift the the shoe for everyday walking from the table using the right arm and place it on the mat", "image": [ "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000176.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_clean_50/episode_010/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_clean_50/episode_010/sample_frames/right_camera/000045.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_clean_50/episode_010/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/place_shoe_aloha-agilex_clean_50/episode_010/sample_frames/right_camera/000000.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift the the shoe for everyday walking from the table using the right arm and place it on the mat`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift the the shoe for everyday walking from the table using the right arm and place it on the mat`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_two_episode_016-train-bf_000234-af_000025-minus-0000038", "task": "the dark inside bowl light blue outside goes at the bottom, the blue bowl rests above.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000311.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000234.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_016/sample_frames/left_camera/000234.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_016/sample_frames/right_camera/000234.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000025.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_016/sample_frames/left_camera/000025.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_randomized_500/episode_016/sample_frames/right_camera/000025.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`the dark inside bowl light blue outside goes at the bottom, the blue bowl rests above.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `the dark inside bowl light blue outside goes at the bottom, the blue bowl rests above.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -88.2, "hop_absolute_value": 88.2, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_three_episode_007-train-bf_000454-af_000263-minus-0000237", "task": "Use arms to stack each the blue bowl carefully.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000483.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000454.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_007/sample_frames/left_camera/000454.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_007/sample_frames/right_camera/000454.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000263.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_007/sample_frames/left_camera/000263.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_007/sample_frames/right_camera/000263.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use arms to stack each the blue bowl carefully.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use arms to stack each the blue bowl carefully.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -35.5, "hop_absolute_value": 35.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_hanging_mug_episode_013-train-bf_000196-af_000125-minus-0000183", "task": "Take the cylindrical mug with curved handle from the table, rotate it, place it in the center, then hang it onto the metallic angular rack with rod pair.", "image": [ "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000331.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000196.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_013/sample_frames/left_camera/000196.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_013/sample_frames/right_camera/000196.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000125.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_013/sample_frames/left_camera/000125.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_013/sample_frames/right_camera/000125.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the cylindrical mug with curved handle from the table, rotate it, place it in the center, then hang it onto the metallic angular rack with rod pair.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the cylindrical mug with curved handle from the table, rotate it, place it in the center, then hang it onto the metallic angular rack with rod pair.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_shake_bottle_episode_010-train-bf_000225-af_000165-minus-0000121", "task": "Use the left arm to grab the bottle with tapered cylindrical shape and shake it.", "image": [ "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_001/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_001/sample_frames/head_camera/000255.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000225.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_010/sample_frames/left_camera/000225.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_010/sample_frames/right_camera/000225.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000165.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_010/sample_frames/left_camera/000165.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_010/sample_frames/right_camera/000165.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the left arm to grab the bottle with tapered cylindrical shape and shake it.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the left arm to grab the bottle with tapered cylindrical shape and shake it.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -26.7, "hop_absolute_value": 26.7, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_bread_basket_episode_002-train-bf_000195-af_000135-minus-0000096", "task": "Shift the golden bread and the medium bread with rounded sides together to the oval basket for bread.", "image": [ "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000227.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000195.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/left_camera/000195.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/right_camera/000195.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000135.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/left_camera/000135.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/right_camera/000135.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Shift the golden bread and the medium bread with rounded sides together to the oval basket for bread.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Shift the golden bread and the medium bread with rounded sides together to the oval basket for bread.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -30.8, "hop_absolute_value": 30.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_018-train-bf_000084-af_000075-minus-0000654", "task": "Place the green-striped cap on yellow bottle, the hand-sized bottle, and the bottle with curved body shape in the rectangular black trash holder using the arms.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000661.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000084.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_018/sample_frames/left_camera/000084.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_018/sample_frames/right_camera/000084.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000075.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_018/sample_frames/left_camera/000075.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_018/sample_frames/right_camera/000075.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the green-striped cap on yellow bottle, the hand-sized bottle, and the bottle with curved body shape in the rectangular black trash holder using the arms.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the green-striped cap on yellow bottle, the hand-sized bottle, and the bottle with curved body shape in the rectangular black trash holder using the arms.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -11.1, "hop_absolute_value": 11.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_003-train-bf_000028-af_000000-minus-0000004", "task": "Pick the medium yellow cylindrical bottle, drop it into the black dustbin, then move the bottle for holding liquids and the rounded bottle with a white neck into the black dustbin.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000662.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_003/sample_frames/head_camera/000028.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_003/sample_frames/left_camera/000028.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_003/sample_frames/right_camera/000028.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_003/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_003/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_clean_50/episode_003/sample_frames/right_camera/000000.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the medium yellow cylindrical bottle, drop it into the black dustbin, then move the bottle for holding liquids and the rounded bottle with a white neck into the black dustbin.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the medium yellow cylindrical bottle, drop it into the black dustbin, then move the bottle for holding liquids and the rounded bottle with a white neck into the black dustbin.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_013-train-bf_000328-af_000205-minus-0000214", "task": "Arrange red block, green block, and blue block from left to right using the right arm, the left arm, and the left arm in the order red, green, blue.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000455.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000328.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_013/sample_frames/left_camera/000328.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_013/sample_frames/right_camera/000328.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000205.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_013/sample_frames/left_camera/000205.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_clean_50/episode_013/sample_frames/right_camera/000205.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Arrange red block, green block, and blue block from left to right using the right arm, the left arm, and the left arm in the order red, green, blue.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Arrange red block, green block, and blue block from left to right using the right arm, the left arm, and the left arm in the order red, green, blue.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -37.5, "hop_absolute_value": 37.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_shake_bottle_episode_018-train-bf_000244-af_000045-minus-0000043", "task": "Use the left arm to pick and shake the bottle with white cap.", "image": [ "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000252.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000244.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_018/sample_frames/left_camera/000244.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_018/sample_frames/right_camera/000244.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_018/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_018/sample_frames/right_camera/000045.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the left arm to pick and shake the bottle with white cap.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the left arm to pick and shake the bottle with white cap.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -82.4, "hop_absolute_value": 82.4, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_three_episode_003-train-bf_000413-af_000305-minus-0000702", "task": "Start with one the bowl with shiny texture, and stack the other two on top.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000463.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_003/sample_frames/head_camera/000413.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_003/sample_frames/left_camera/000413.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_003/sample_frames/right_camera/000413.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_003/sample_frames/head_camera/000305.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_003/sample_frames/left_camera/000305.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_003/sample_frames/right_camera/000305.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Start with one the bowl with shiny texture, and stack the other two on top.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Start with one the bowl with shiny texture, and stack the other two on top.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -23.3, "hop_absolute_value": 23.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_laptop_episode_027-train-bf_000180-af_000030-minus-0000045", "task": "Raise the black laptop with Saturn image on screen using the right arm.", "image": [ "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000225.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000180.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000180.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000180.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000030.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000030.png", "images/robotwin2_agilex_part1/open_laptop_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000030.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Raise the black laptop with Saturn image on screen using the right arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Raise the black laptop with Saturn image on screen using the right arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_can_basket_episode_004-train-bf_000200-af_000000-minus-0000014", "task": "Lift the red beverage can with curved top, set it in the light brown oval basket, and elevate the light brown oval basket.", "image": [ "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000254.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000200.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/left_camera/000200.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/right_camera/000200.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_004/sample_frames/right_camera/000000.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift the red beverage can with curved top, set it in the light brown oval basket, and elevate the light brown oval basket.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift the red beverage can with curved top, set it in the light brown oval basket, and elevate the light brown oval basket.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_028-train-bf_000093-af_000037-minus-0000176", "task": "Carefully grab the yellow bottle body with tapered neck, the clear bottle with blue sections, and the black and red bottle and position them in the trash bin.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000664.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000093.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_028/sample_frames/left_camera/000093.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_028/sample_frames/right_camera/000093.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_028/sample_frames/head_camera/000037.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_028/sample_frames/left_camera/000037.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_028/sample_frames/right_camera/000037.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Carefully grab the yellow bottle body with tapered neck, the clear bottle with blue sections, and the black and red bottle and position them in the trash bin.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Carefully grab the yellow bottle body with tapered neck, the clear bottle with blue sections, and the black and red bottle and position them in the trash bin.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_026-train-bf_000286-af_000232-minus-0000647", "task": "Position red block, green block, and blue block in a left-to-right sequence, forming a row.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000458.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_026/sample_frames/head_camera/000286.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_026/sample_frames/left_camera/000286.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_026/sample_frames/right_camera/000286.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_026/sample_frames/head_camera/000232.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_026/sample_frames/left_camera/000232.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_026/sample_frames/right_camera/000232.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Position red block, green block, and blue block in a left-to-right sequence, forming a row.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Position red block, green block, and blue block in a left-to-right sequence, forming a row.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -19.0, "hop_absolute_value": 19.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_006-train-bf_000433-af_000273-minus-0000239", "task": "Using the right arm, the left arm, and the right arm, arrange red block, green block, and blue block sequentially from left to right in a single row.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000457.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000433.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000433.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000433.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000273.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000273.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000273.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Using the right arm, the left arm, and the right arm, arrange red block, green block, and blue block sequentially from left to right in a single row.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Using the right arm, the left arm, and the right arm, arrange red block, green block, and blue block sequentially from left to right in a single row.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -35.5, "hop_absolute_value": 35.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_turn_switch_episode_005-train-bf_000075-af_000060-minus-0000026", "task": "Switch the beige switch to its active state", "image": [ "images/robotwin2_agilex_part1/turn_switch_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/turn_switch_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000097.png", "images/robotwin2_agilex_part1/turn_switch_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000075.png", "images/robotwin2_agilex_part1/turn_switch_aloha-agilex_randomized_500/episode_005/sample_frames/left_camera/000075.png", "images/robotwin2_agilex_part1/turn_switch_aloha-agilex_randomized_500/episode_005/sample_frames/right_camera/000075.png", "images/robotwin2_agilex_part1/turn_switch_aloha-agilex_randomized_500/episode_005/sample_frames/head_camera/000060.png", "images/robotwin2_agilex_part1/turn_switch_aloha-agilex_randomized_500/episode_005/sample_frames/left_camera/000060.png", "images/robotwin2_agilex_part1/turn_switch_aloha-agilex_randomized_500/episode_005/sample_frames/right_camera/000060.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Switch the beige switch to its active state`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Switch the beige switch to its active state`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_three_episode_006-train-bf_000140-af_000089-minus-0000210", "task": "Place red block, green block, and blue block in the center, then arrange green block on red block and blue block on green block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000459.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000140.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000140.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000140.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000089.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000089.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000089.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place red block, green block, and blue block in the center, then arrange green block on red block and blue block on green block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place red block, green block, and blue block in the center, then arrange green block on red block and blue block on green block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -36.4, "hop_absolute_value": 36.4, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_three_episode_002-train-bf_000340-af_000276-minus-0000259", "task": "Place the first the small bowl with brown edge down, then stack the remaining two.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_020/sample_frames/head_camera/000488.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000340.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000340.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000340.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000276.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000276.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000276.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the first the small bowl with brown edge down, then stack the remaining two.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the first the small bowl with brown edge down, then stack the remaining two.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -16.0, "hop_absolute_value": 16.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_can_basket_episode_016-train-bf_000071-af_000036-minus-0000092", "task": "Pick up the drink can with smooth red finish, put it into the bright yellow basket, and lift the bright yellow basket.", "image": [ "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000247.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000071.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_016/sample_frames/left_camera/000071.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_016/sample_frames/right_camera/000071.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000036.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_016/sample_frames/left_camera/000036.png", "images/robotwin2_agilex_part1/place_can_basket_aloha-agilex_clean_50/episode_016/sample_frames/right_camera/000036.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the drink can with smooth red finish, put it into the bright yellow basket, and lift the bright yellow basket.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the drink can with smooth red finish, put it into the bright yellow basket, and lift the bright yellow basket.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_block_episode_016-train-bf_000233-af_000129-minus-0000132", "task": "Hand the red block to the right arm.", "image": [ "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_017/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_017/sample_frames/head_camera/000284.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000233.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_016/sample_frames/left_camera/000233.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_016/sample_frames/right_camera/000233.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000129.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_016/sample_frames/left_camera/000129.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_016/sample_frames/right_camera/000129.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Hand the red block to the right arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Hand the red block to the right arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -47.1, "hop_absolute_value": 47.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_010-train-bf_000273-af_000205-minus-0000702", "task": "Use the right arm, the left arm, and the left arm to arrange large block, medium block, small block in order here.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000462.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000273.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_010/sample_frames/left_camera/000273.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_010/sample_frames/right_camera/000273.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000205.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_010/sample_frames/left_camera/000205.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_010/sample_frames/right_camera/000205.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the right arm, the left arm, and the left arm to arrange large block, medium block, small block in order here.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the right arm, the left arm, and the left arm to arrange large block, medium block, small block in order here.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_bread_basket_episode_002-train-bf_000165-af_000135-minus-0000115", "task": "Shift the golden bread and the medium bread with rounded sides together to the oval basket for bread.", "image": [ "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000226.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000165.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/left_camera/000165.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/right_camera/000165.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/head_camera/000135.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/left_camera/000135.png", "images/robotwin2_agilex_part1/place_bread_basket_aloha-agilex_clean_50/episode_002/sample_frames/right_camera/000135.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Shift the golden bread and the medium bread with rounded sides together to the oval basket for bread.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Shift the golden bread and the medium bread with rounded sides together to the oval basket for bread.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -18.2, "hop_absolute_value": 18.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_bread_skillet_episode_007-train-bf_000149-af_000120-minus-0000055", "task": "Use the left arm to pick the small golden brown bread and place it into the medium skillet.", "image": [ "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000163.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000149.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_007/sample_frames/left_camera/000149.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_007/sample_frames/right_camera/000149.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000120.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_007/sample_frames/left_camera/000120.png", "images/robotwin2_agilex_part1/place_bread_skillet_aloha-agilex_randomized_500/episode_007/sample_frames/right_camera/000120.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the left arm to pick the small golden brown bread and place it into the medium skillet.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the left arm to pick the small golden brown bread and place it into the medium skillet.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -40.0, "hop_absolute_value": 40.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_hanging_mug_episode_010-train-bf_000100-af_000038-minus-0000102", "task": "Take the drinking mug shiny black exterior, rotate it, place it in the center, then hang it on the rack with two rods sticking out.", "image": [ "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000336.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000100.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/left_camera/000100.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/right_camera/000100.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000038.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/left_camera/000038.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_randomized_500/episode_010/sample_frames/right_camera/000038.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the drinking mug shiny black exterior, rotate it, place it in the center, then hang it on the rack with two rods sticking out.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the drinking mug shiny black exterior, rotate it, place it in the center, then hang it on the rack with two rods sticking out.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_000-train-bf_000150-af_000030-minus-0000070", "task": "Use the left arm to pull the gray and yellow microwave with rounded edges open.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_008/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_008/sample_frames/head_camera/000459.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000150.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_000/sample_frames/left_camera/000150.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_000/sample_frames/right_camera/000150.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000030.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_000/sample_frames/left_camera/000030.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_000/sample_frames/right_camera/000030.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the left arm to pull the gray and yellow microwave with rounded edges open.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the left arm to pull the gray and yellow microwave with rounded edges open.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_block_episode_010-train-bf_000157-af_000100-minus-0000138", "task": "Lift the red block from the table with the left arm, pass it to the right arm, and place it on the blue pad.", "image": [ "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000288.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000157.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_010/sample_frames/left_camera/000157.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_010/sample_frames/right_camera/000157.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000100.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_010/sample_frames/left_camera/000100.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_010/sample_frames/right_camera/000100.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift the red block from the table with the left arm, pass it to the right arm, and place it on the blue pad.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift the red block from the table with the left arm, pass it to the right arm, and place it on the blue pad.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -36.4, "hop_absolute_value": 36.4, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_cans_plasticbox_episode_006-train-bf_000177-af_000045-minus-0000062", "task": "First pick the red can with white Coca-Cola logo, place it in the medium storage plasticbox, then repeat for the compact metal drink can.", "image": [ "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000285.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000177.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000177.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000177.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_006/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_006/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/place_cans_plasticbox_aloha-agilex_randomized_500/episode_006/sample_frames/right_camera/000045.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`First pick the red can with white Coca-Cola logo, place it in the medium storage plasticbox, then repeat for the compact metal drink can.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `First pick the red can with white Coca-Cola logo, place it in the medium storage plasticbox, then repeat for the compact metal drink can.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.0, "hop_absolute_value": 75.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_block_episode_007-train-bf_000171-af_000029-minus-0000049", "task": "Pick up the red block using the left arm, give it to the right arm, and place it on the blue pad.", "image": [ "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000276.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000171.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_007/sample_frames/left_camera/000171.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_007/sample_frames/right_camera/000171.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_007/sample_frames/head_camera/000029.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_007/sample_frames/left_camera/000029.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_007/sample_frames/right_camera/000029.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the red block using the left arm, give it to the right arm, and place it on the blue pad.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the red block using the left arm, give it to the right arm, and place it on the blue pad.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -83.3, "hop_absolute_value": 83.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_size_episode_016-train-bf_000055-af_000000-minus-0000001", "task": "Use the left arm, the right arm, and the left arm to center large block, medium block, and small block in order.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000473.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000055.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_016/sample_frames/left_camera/000055.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_016/sample_frames/right_camera/000055.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_016/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_size_aloha-agilex_randomized_500/episode_016/sample_frames/right_camera/000000.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the left arm, the right arm, and the left arm to center large block, medium block, and small block in order.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the left arm, the right arm, and the left arm to center large block, medium block, and small block in order.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_dual_shoes_episode_000-train-bf_000218-af_000140-minus-0000071", "task": "Stick both the plastic brown shoe in the rectangular orange shoe-box, shoe tips pointing left.", "image": [ "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000215.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000218.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_000/sample_frames/left_camera/000218.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_000/sample_frames/right_camera/000218.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000140.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_000/sample_frames/left_camera/000140.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_000/sample_frames/right_camera/000140.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Stick both the plastic brown shoe in the rectangular orange shoe-box, shoe tips pointing left.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Stick both the plastic brown shoe in the rectangular orange shoe-box, shoe tips pointing left.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -56.2, "hop_absolute_value": 56.2, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_move_playingcard_away_episode_012-train-bf_000075-af_000015-minus-0000010", "task": "Lift the small playingcards container and shift it outward.", "image": [ "images/robotwin2_agilex_part1/move_playingcard_away_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/move_playingcard_away_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000121.png", "images/robotwin2_agilex_part1/move_playingcard_away_aloha-agilex_randomized_500/episode_012/sample_frames/head_camera/000075.png", "images/robotwin2_agilex_part1/move_playingcard_away_aloha-agilex_randomized_500/episode_012/sample_frames/left_camera/000075.png", "images/robotwin2_agilex_part1/move_playingcard_away_aloha-agilex_randomized_500/episode_012/sample_frames/right_camera/000075.png", "images/robotwin2_agilex_part1/move_playingcard_away_aloha-agilex_randomized_500/episode_012/sample_frames/head_camera/000015.png", "images/robotwin2_agilex_part1/move_playingcard_away_aloha-agilex_randomized_500/episode_012/sample_frames/left_camera/000015.png", "images/robotwin2_agilex_part1/move_playingcard_away_aloha-agilex_randomized_500/episode_012/sample_frames/right_camera/000015.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift the small playingcards container and shift it outward.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift the small playingcards container and shift it outward.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -80.0, "hop_absolute_value": 80.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_dual_shoes_episode_022-train-bf_000225-af_000184-minus-0000104", "task": "Lift two the white and green shoe, ensure left tips, drop in the shoe-box with Nike printed logo.", "image": [ "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000230.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000225.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_022/sample_frames/left_camera/000225.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_022/sample_frames/right_camera/000225.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000184.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_022/sample_frames/left_camera/000184.png", "images/robotwin2_agilex_part1/place_dual_shoes_aloha-agilex_randomized_500/episode_022/sample_frames/right_camera/000184.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift two the white and green shoe, ensure left tips, drop in the shoe-box with Nike printed logo.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift two the white and green shoe, ensure left tips, drop in the shoe-box with Nike printed logo.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -31.2, "hop_absolute_value": 31.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_three_episode_002-train-bf_000408-af_000089-minus-0000121", "task": "Place the first the small bowl with brown edge down, then stack the remaining two.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_008/sample_frames/head_camera/000451.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000408.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000408.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000408.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000089.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_002/sample_frames/left_camera/000089.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_002/sample_frames/right_camera/000089.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place the first the small bowl with brown edge down, then stack the remaining two.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place the first the small bowl with brown edge down, then stack the remaining two.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -75.9, "hop_absolute_value": 75.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_bottles_dustbin_episode_026-train-bf_000140-af_000070-minus-0000188", "task": "Pick the smooth yellow bottle with ridged cap, drop it into the medium black garbage container, then move the blue-accented cylindrical bottle and the bottle with curved body shape into the medium black garbage container.", "image": [ "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000659.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_026/sample_frames/head_camera/000140.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_026/sample_frames/left_camera/000140.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_026/sample_frames/right_camera/000140.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_026/sample_frames/head_camera/000070.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_026/sample_frames/left_camera/000070.png", "images/robotwin2_agilex_part1/put_bottles_dustbin_aloha-agilex_randomized_500/episode_026/sample_frames/right_camera/000070.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the smooth yellow bottle with ridged cap, drop it into the medium black garbage container, then move the blue-accented cylindrical bottle and the bottle with curved body shape into the medium black garbage container.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the smooth yellow bottle with ridged cap, drop it into the medium black garbage container, then move the blue-accented cylindrical bottle and the bottle with curved body shape into the medium black garbage container.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_move_can_pot_episode_016-train-bf_000148-af_000060-minus-0000029", "task": "Use the right arm to grab the red cylindrical sauce can and transfer it near the kitchenpot with flat lid and knob", "image": [ "images/robotwin2_agilex_part1/move_can_pot_aloha-agilex_randomized_500/episode_029/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/move_can_pot_aloha-agilex_randomized_500/episode_029/sample_frames/head_camera/000155.png", "images/robotwin2_agilex_part1/move_can_pot_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000148.png", "images/robotwin2_agilex_part1/move_can_pot_aloha-agilex_randomized_500/episode_016/sample_frames/left_camera/000148.png", "images/robotwin2_agilex_part1/move_can_pot_aloha-agilex_randomized_500/episode_016/sample_frames/right_camera/000148.png", "images/robotwin2_agilex_part1/move_can_pot_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000060.png", "images/robotwin2_agilex_part1/move_can_pot_aloha-agilex_randomized_500/episode_016/sample_frames/left_camera/000060.png", "images/robotwin2_agilex_part1/move_can_pot_aloha-agilex_randomized_500/episode_016/sample_frames/right_camera/000060.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the right arm to grab the red cylindrical sauce can and transfer it near the kitchenpot with flat lid and knob`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the right arm to grab the red cylindrical sauce can and transfer it near the kitchenpot with flat lid and knob`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_three_episode_011-train-bf_000334-af_000263-minus-0000655", "task": "Stack the blue bowl with slightly rounded base, starting with the largest bowl at the bottom.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000479.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000334.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000334.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000334.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000263.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000263.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000263.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Stack the blue bowl with slightly rounded base, starting with the largest bowl at the bottom.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Stack the blue bowl with slightly rounded base, starting with the largest bowl at the bottom.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_press_stapler_episode_016-train-bf_000045-af_000000-minus-0000003", "task": "Press the black stapler with rectangular base downward using your arm the right arm.", "image": [ "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_clean_50/episode_017/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_clean_50/episode_017/sample_frames/head_camera/000122.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_clean_50/episode_016/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_clean_50/episode_016/sample_frames/right_camera/000045.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_clean_50/episode_016/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_clean_50/episode_016/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/press_stapler_aloha-agilex_clean_50/episode_016/sample_frames/right_camera/000000.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Press the black stapler with rectangular base downward using your arm the right arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Press the black stapler with rectangular base downward using your arm the right arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_put_object_cabinet_episode_013-train-bf_000241-af_000084-minus-0000060", "task": "Open the drawer of the medium-sized cabinet and place the rectangular glossy phone with dark color into it.", "image": [ "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000266.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000241.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_013/sample_frames/left_camera/000241.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_013/sample_frames/right_camera/000241.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_013/sample_frames/head_camera/000084.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_013/sample_frames/left_camera/000084.png", "images/robotwin2_agilex_part1/put_object_cabinet_aloha-agilex_clean_50/episode_013/sample_frames/right_camera/000084.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Open the drawer of the medium-sized cabinet and place the rectangular glossy phone with dark color into it.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Open the drawer of the medium-sized cabinet and place the rectangular glossy phone with dark color into it.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.3, "hop_absolute_value": 73.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_three_episode_011-train-bf_000276-af_000263-minus-0000587", "task": "Stack the blue bowl with slightly rounded base, starting with the largest bowl at the bottom.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000463.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000276.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000276.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000276.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000263.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000263.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000263.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Stack the blue bowl with slightly rounded base, starting with the largest bowl at the bottom.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Stack the blue bowl with slightly rounded base, starting with the largest bowl at the bottom.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -4.8, "hop_absolute_value": 4.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_shake_bottle_horizontally_episode_015-train-bf_000240-af_000225-minus-0000170", "task": "Pick up the light and medium-sized bottle and shake it from side to side.", "image": [ "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000273.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000240.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_015/sample_frames/left_camera/000240.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_015/sample_frames/right_camera/000240.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000225.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_015/sample_frames/left_camera/000225.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_clean_50/episode_015/sample_frames/right_camera/000225.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the light and medium-sized bottle and shake it from side to side.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the light and medium-sized bottle and shake it from side to side.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -6.2, "hop_absolute_value": 6.2, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_pick_dual_bottles_episode_006-train-bf_000105-af_000045-minus-0000023", "task": "Take the narrow-necked bottle with one arm, hold the medium-sized green bottle too.", "image": [ "images/robotwin2_agilex_part1/pick_dual_bottles_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/pick_dual_bottles_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000123.png", "images/robotwin2_agilex_part1/pick_dual_bottles_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000105.png", "images/robotwin2_agilex_part1/pick_dual_bottles_aloha-agilex_clean_50/episode_006/sample_frames/left_camera/000105.png", "images/robotwin2_agilex_part1/pick_dual_bottles_aloha-agilex_clean_50/episode_006/sample_frames/right_camera/000105.png", "images/robotwin2_agilex_part1/pick_dual_bottles_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/pick_dual_bottles_aloha-agilex_clean_50/episode_006/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/pick_dual_bottles_aloha-agilex_clean_50/episode_006/sample_frames/right_camera/000045.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the narrow-necked bottle with one arm, hold the medium-sized green bottle too.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the narrow-necked bottle with one arm, hold the medium-sized green bottle too.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -57.1, "hop_absolute_value": 57.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_009-train-bf_000245-af_000095-minus-0000137", "task": "First, place red block on the far left, then green block in the middle, and finish with blue block on the right.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000451.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000245.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_009/sample_frames/left_camera/000245.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_009/sample_frames/right_camera/000245.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000095.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_009/sample_frames/left_camera/000095.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_009/sample_frames/right_camera/000095.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`First, place red block on the far left, then green block in the middle, and finish with blue block on the right.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `First, place red block on the far left, then green block in the middle, and finish with blue block on the right.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -61.1, "hop_absolute_value": 61.1, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_021-train-bf_000356-af_000082-minus-0000092", "task": "Start with red block, followed by green block and blue block, placing them in order left to right.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_001/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_001/sample_frames/head_camera/000463.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000356.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_021/sample_frames/left_camera/000356.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_021/sample_frames/right_camera/000356.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_021/sample_frames/head_camera/000082.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_021/sample_frames/left_camera/000082.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_021/sample_frames/right_camera/000082.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Start with red block, followed by green block and blue block, placing them in order left to right.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Start with red block, followed by green block and blue block, placing them in order left to right.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -76.9, "hop_absolute_value": 76.9, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_block_episode_018-train-bf_000157-af_000114-minus-0000169", "task": "Hand the red block to the right arm.", "image": [ "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_011/sample_frames/head_camera/000279.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000157.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_018/sample_frames/left_camera/000157.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_018/sample_frames/right_camera/000157.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_018/sample_frames/head_camera/000114.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_018/sample_frames/left_camera/000114.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_randomized_500/episode_018/sample_frames/right_camera/000114.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Hand the red block to the right arm.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Hand the red block to the right arm.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -27.3, "hop_absolute_value": 27.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_three_episode_004-train-bf_000195-af_000051-minus-0000110", "task": "Move red block, green block, and blue block to the middle first, then stack blue block on green block and green block on red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000480.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000195.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_004/sample_frames/left_camera/000195.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_004/sample_frames/right_camera/000195.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_004/sample_frames/head_camera/000051.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_004/sample_frames/left_camera/000051.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_004/sample_frames/right_camera/000051.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Move red block, green block, and blue block to the middle first, then stack blue block on green block and green block on red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Move red block, green block, and blue block to the middle first, then stack blue block on green block and green block on red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -73.3, "hop_absolute_value": 73.3, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_027-train-bf_000273-af_000205-minus-0000726", "task": "Position red block first, followed by green block, and end with blue block in a row from left to right.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000456.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000273.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000273.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000273.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000205.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000205.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000205.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Position red block first, followed by green block, and end with blue block in a row from left to right.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Position red block first, followed by green block, and end with blue block in a row from left to right.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_move_pillbottle_pad_episode_014-train-bf_000045-af_000030-minus-0000042", "task": "Set the round bottle with orange screw-top down on the pad after picking it.", "image": [ "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_001/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_001/sample_frames/head_camera/000148.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_014/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_014/sample_frames/right_camera/000045.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000030.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_014/sample_frames/left_camera/000030.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_014/sample_frames/right_camera/000030.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Set the round bottle with orange screw-top down on the pad after picking it.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Set the round bottle with orange screw-top down on the pad after picking it.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -33.3, "hop_absolute_value": 33.3, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_016-train-bf_000454-af_000315-minus-0000786", "task": "Place red block first, set green block beside it, and complete the row by positioning blue block.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000460.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000454.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_016/sample_frames/left_camera/000454.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_016/sample_frames/right_camera/000454.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000315.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_016/sample_frames/left_camera/000315.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_016/sample_frames/right_camera/000315.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Place red block first, set green block beside it, and complete the row by positioning blue block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Place red block first, set green block beside it, and complete the row by positioning blue block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -28.1, "hop_absolute_value": 28.1, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_mouse_pad_episode_014-train-bf_000135-af_000000-minus-0000008", "task": "Pick up the the palm-fitting mouse with the right arm and place it on the Red mat.", "image": [ "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000161.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000135.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_014/sample_frames/left_camera/000135.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_014/sample_frames/right_camera/000135.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_014/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_014/sample_frames/right_camera/000000.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the the palm-fitting mouse with the right arm and place it on the Red mat.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the the palm-fitting mouse with the right arm and place it on the Red mat.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_move_pillbottle_pad_episode_015-train-bf_000120-af_000045-minus-0000024", "task": "Pick the white cap on orange bottle with the right arm and drop it onto the pad.", "image": [ "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000155.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000120.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_015/sample_frames/left_camera/000120.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_015/sample_frames/right_camera/000120.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_015/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/move_pillbottle_pad_aloha-agilex_clean_50/episode_015/sample_frames/right_camera/000045.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the white cap on orange bottle with the right arm and drop it onto the pad.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the white cap on orange bottle with the right arm and drop it onto the pad.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_shake_bottle_episode_022-train-bf_000180-af_000165-minus-0000140", "task": "Pick the yellow bottle and perform shaking motion.", "image": [ "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_002/sample_frames/head_camera/000256.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000180.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_022/sample_frames/left_camera/000180.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_022/sample_frames/right_camera/000180.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_022/sample_frames/head_camera/000165.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_022/sample_frames/left_camera/000165.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_022/sample_frames/right_camera/000165.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the yellow bottle and perform shaking motion.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the yellow bottle and perform shaking motion.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -8.3, "hop_absolute_value": 8.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_handover_block_episode_012-train-bf_000223-af_000029-minus-0000039", "task": "Use left arm to grab the block and hand it to the right.", "image": [ "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_008/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_008/sample_frames/head_camera/000280.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_012/sample_frames/head_camera/000223.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_012/sample_frames/left_camera/000223.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_012/sample_frames/right_camera/000223.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_012/sample_frames/head_camera/000029.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_012/sample_frames/left_camera/000029.png", "images/robotwin2_agilex_part1/handover_block_aloha-agilex_clean_50/episode_012/sample_frames/right_camera/000029.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use left arm to grab the block and hand it to the right.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use left arm to grab the block and hand it to the right.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -87.5, "hop_absolute_value": 87.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_two_episode_015-train-bf_000127-af_000051-minus-0000109", "task": "Relocate red block and green block to the table's midpoint, and stack green block on top of red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_018/sample_frames/head_camera/000299.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000127.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_015/sample_frames/left_camera/000127.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_015/sample_frames/right_camera/000127.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_015/sample_frames/head_camera/000051.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_015/sample_frames/left_camera/000051.png", "images/robotwin2_agilex_part1/stack_blocks_two_aloha-agilex_clean_50/episode_015/sample_frames/right_camera/000051.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Relocate red block and green block to the table's midpoint, and stack green block on top of red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Relocate red block and green block to the table's midpoint, and stack green block on top of red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -60.0, "hop_absolute_value": 60.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_dump_bin_bigbin_episode_014-train-bf_000240-af_000180-minus-0000181", "task": "Use the arm to empty the compact tabletop trashbin into the bin.", "image": [ "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000325.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000240.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_014/sample_frames/left_camera/000240.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_014/sample_frames/right_camera/000240.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_014/sample_frames/head_camera/000180.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_014/sample_frames/left_camera/000180.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_clean_50/episode_014/sample_frames/right_camera/000180.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Use the arm to empty the compact tabletop trashbin into the bin.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Use the arm to empty the compact tabletop trashbin into the bin.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.0, "hop_absolute_value": 25.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_019-train-bf_000150-af_000120-minus-0000634", "task": "Pull the handle to open the microwave with textured top panel.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_012/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_012/sample_frames/head_camera/000457.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000150.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000150.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000150.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_019/sample_frames/head_camera/000120.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_019/sample_frames/left_camera/000120.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_clean_50/episode_019/sample_frames/right_camera/000120.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pull the handle to open the microwave with textured top panel.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pull the handle to open the microwave with textured top panel.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_open_microwave_episode_019-train-bf_000360-af_000285-minus-0000863", "task": "Access the inside by opening the dark gray kitchen microwave.", "image": [ "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000490.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000360.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_019/sample_frames/left_camera/000360.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_019/sample_frames/right_camera/000360.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_019/sample_frames/head_camera/000285.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_019/sample_frames/left_camera/000285.png", "images/robotwin2_agilex_part1/open_microwave_aloha-agilex_randomized_500/episode_019/sample_frames/right_camera/000285.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Access the inside by opening the dark gray kitchen microwave.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Access the inside by opening the dark gray kitchen microwave.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.8, "hop_absolute_value": 20.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_three_episode_006-train-bf_000127-af_000064-minus-0000184", "task": "With the right arm, the right arm, and the left arm, shift red block, green block, and blue block to the center and arrange blue block over green block, green block on red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_009/sample_frames/head_camera/000472.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000127.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_006/sample_frames/left_camera/000127.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_006/sample_frames/right_camera/000127.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_006/sample_frames/head_camera/000064.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_006/sample_frames/left_camera/000064.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_clean_50/episode_006/sample_frames/right_camera/000064.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`With the right arm, the right arm, and the left arm, shift red block, green block, and blue block to the center and arrange blue block over green block, green block on red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `With the right arm, the right arm, and the left arm, shift red block, green block, and blue block to the center and arrange blue block over green block, green block on red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_hanging_mug_episode_011-train-bf_000075-af_000000-minus-0000003", "task": "Pick up the glossy beige ceramic drinking mug, turn it, place it centrally, then hang the glossy beige ceramic drinking mug onto the dark gray rack.", "image": [ "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_000/sample_frames/head_camera/000337.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000075.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000075.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000075.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_011/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/hanging_mug_aloha-agilex_clean_50/episode_011/sample_frames/right_camera/000000.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the glossy beige ceramic drinking mug, turn it, place it centrally, then hang the glossy beige ceramic drinking mug onto the dark gray rack.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the glossy beige ceramic drinking mug, turn it, place it centrally, then hang the glossy beige ceramic drinking mug onto the dark gray rack.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stamp_seal_episode_016-train-bf_000120-af_000045-minus-0000024", "task": "Grab the round yellow seal base and press Tan", "image": [ "images/robotwin2_agilex_part1/stamp_seal_aloha-agilex_randomized_500/episode_011/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stamp_seal_aloha-agilex_randomized_500/episode_011/sample_frames/head_camera/000147.png", "images/robotwin2_agilex_part1/stamp_seal_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000120.png", "images/robotwin2_agilex_part1/stamp_seal_aloha-agilex_randomized_500/episode_016/sample_frames/left_camera/000120.png", "images/robotwin2_agilex_part1/stamp_seal_aloha-agilex_randomized_500/episode_016/sample_frames/right_camera/000120.png", "images/robotwin2_agilex_part1/stamp_seal_aloha-agilex_randomized_500/episode_016/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/stamp_seal_aloha-agilex_randomized_500/episode_016/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/stamp_seal_aloha-agilex_randomized_500/episode_016/sample_frames/right_camera/000045.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Grab the round yellow seal base and press Tan`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Grab the round yellow seal base and press Tan`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -62.5, "hop_absolute_value": 62.5, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_027-train-bf_000357-af_000041-minus-0000055", "task": "Position red block first, followed by green block, and end with blue block in a row from left to right.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_024/sample_frames/head_camera/000453.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000357.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000357.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000357.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000041.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000041.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000041.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Position red block first, followed by green block, and end with blue block in a row from left to right.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Position red block first, followed by green block, and end with blue block in a row from left to right.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -88.5, "hop_absolute_value": 88.5, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_blocks_ranking_rgb_episode_004-train-bf_000370-af_000273-minus-0000721", "task": "Position red block, green block, and blue block on a surface from left to right in the order red, green, blue.", "image": [ "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000451.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_004/sample_frames/head_camera/000370.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_004/sample_frames/left_camera/000370.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_004/sample_frames/right_camera/000370.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_004/sample_frames/head_camera/000273.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_004/sample_frames/left_camera/000273.png", "images/robotwin2_agilex_part1/blocks_ranking_rgb_aloha-agilex_randomized_500/episode_004/sample_frames/right_camera/000273.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Position red block, green block, and blue block on a surface from left to right in the order red, green, blue.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Position red block, green block, and blue block on a surface from left to right in the order red, green, blue.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -25.9, "hop_absolute_value": 25.9, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_dump_bin_bigbin_episode_015-train-bf_000300-af_000165-minus-0000530", "task": "Secure the light gray waste container and pour the balls.", "image": [ "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_000/sample_frames/head_camera/000332.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_015/sample_frames/head_camera/000300.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_015/sample_frames/left_camera/000300.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_015/sample_frames/right_camera/000300.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_015/sample_frames/head_camera/000165.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_015/sample_frames/left_camera/000165.png", "images/robotwin2_agilex_part1/dump_bin_bigbin_aloha-agilex_randomized_500/episode_015/sample_frames/right_camera/000165.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Secure the light gray waste container and pour the balls.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Secure the light gray waste container and pour the balls.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -45.0, "hop_absolute_value": 45.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_blocks_three_episode_013-train-bf_000167-af_000000-minus-0000015", "task": "Shift red block, green block, and blue block to the center using the right arm, the left arm, the right arm, then pile blue block on green block, green block on red block.", "image": [ "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_009/sample_frames/head_camera/000476.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000167.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_013/sample_frames/left_camera/000167.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_013/sample_frames/right_camera/000167.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_013/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_013/sample_frames/left_camera/000000.png", "images/robotwin2_agilex_part1/stack_blocks_three_aloha-agilex_randomized_500/episode_013/sample_frames/right_camera/000000.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Shift red block, green block, and blue block to the center using the right arm, the left arm, the right arm, then pile blue block on green block, green block on red block.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Shift red block, green block, and blue block to the center using the right arm, the left arm, the right arm, then pile blue block on green block, green block on red block.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -100.0, "hop_absolute_value": 100.0, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_mouse_pad_episode_027-train-bf_000045-af_000015-minus-0000022", "task": "Pick up the the plastic mouse with smooth finish with the left arm and place it on the Black mat.", "image": [ "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_004/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_004/sample_frames/head_camera/000157.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000045.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000045.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000045.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_027/sample_frames/head_camera/000015.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_027/sample_frames/left_camera/000015.png", "images/robotwin2_agilex_part1/place_mouse_pad_aloha-agilex_randomized_500/episode_027/sample_frames/right_camera/000015.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the the plastic mouse with smooth finish with the left arm and place it on the Black mat.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the the plastic mouse with smooth finish with the left arm and place it on the Black mat.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -66.7, "hop_absolute_value": 66.7, "hop_category": { "absolute_category": "large", "direction": "regression", "combined_category": "regression_large" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_three_episode_026-train-bf_000195-af_000167-minus-0000260", "task": "Pick up the ceramic bowl with earthy brown trim and stack them from biggest to smallest.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_025/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_025/sample_frames/head_camera/000470.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_026/sample_frames/head_camera/000195.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_026/sample_frames/left_camera/000195.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_026/sample_frames/right_camera/000195.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_026/sample_frames/head_camera/000167.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_026/sample_frames/left_camera/000167.png", "images/robotwin2_agilex_part1/stack_bowls_three_aloha-agilex_randomized_500/episode_026/sample_frames/right_camera/000167.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick up the ceramic bowl with earthy brown trim and stack them from biggest to smallest.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick up the ceramic bowl with earthy brown trim and stack them from biggest to smallest.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -13.3, "hop_absolute_value": 13.3, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_stack_bowls_two_episode_008-train-bf_000172-af_000084-minus-0000119", "task": "the medium-sized bowl for holding food goes at the bottom, the small bowl with brown edge rests above.", "image": [ "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_012/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_012/sample_frames/head_camera/000302.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_008/sample_frames/head_camera/000172.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_008/sample_frames/left_camera/000172.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_008/sample_frames/right_camera/000172.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_008/sample_frames/head_camera/000084.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_008/sample_frames/left_camera/000084.png", "images/robotwin2_agilex_part1/stack_bowls_two_aloha-agilex_clean_50/episode_008/sample_frames/right_camera/000084.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`the medium-sized bowl for holding food goes at the bottom, the small bowl with brown edge rests above.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `the medium-sized bowl for holding food goes at the bottom, the small bowl with brown edge rests above.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_shake_bottle_episode_010-train-bf_000210-af_000135-minus-0000116", "task": "Pick the black bottle up with the left arm and shake.", "image": [ "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_003/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_003/sample_frames/head_camera/000254.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000210.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_010/sample_frames/left_camera/000210.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_010/sample_frames/right_camera/000210.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_010/sample_frames/head_camera/000135.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_010/sample_frames/left_camera/000135.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_clean_50/episode_010/sample_frames/right_camera/000135.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the black bottle up with the left arm and shake.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the black bottle up with the left arm and shake.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -35.7, "hop_absolute_value": 35.7, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_shake_bottle_episode_017-train-bf_000240-af_000120-minus-0000091", "task": "Pick the plastic drink bottle up with the left arm and shake.", "image": [ "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000254.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000240.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_017/sample_frames/left_camera/000240.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_017/sample_frames/right_camera/000240.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_017/sample_frames/head_camera/000120.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_017/sample_frames/left_camera/000120.png", "images/robotwin2_agilex_part1/shake_bottle_aloha-agilex_randomized_500/episode_017/sample_frames/right_camera/000120.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Pick the plastic drink bottle up with the left arm and shake.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Pick the plastic drink bottle up with the left arm and shake.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -50.0, "hop_absolute_value": 50.0, "hop_category": { "absolute_category": "medium", "direction": "regression", "combined_category": "regression_medium" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_shake_bottle_horizontally_episode_011-train-bf_000240-af_000195-minus-0000155", "task": "Lift the bottle with blue cap and shake it horizontally.", "image": [ "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_010/sample_frames/head_camera/000278.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_011/sample_frames/head_camera/000240.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_011/sample_frames/left_camera/000240.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_011/sample_frames/right_camera/000240.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_011/sample_frames/head_camera/000195.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_011/sample_frames/left_camera/000195.png", "images/robotwin2_agilex_part1/shake_bottle_horizontally_aloha-agilex_randomized_500/episode_011/sample_frames/right_camera/000195.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Lift the bottle with blue cap and shake it horizontally.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Lift the bottle with blue cap and shake it horizontally.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -18.8, "hop_absolute_value": 18.8, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" }, { "id": "robotwin_place_phone_stand_episode_015-train-bf_000075-af_000060-minus-0000040", "task": "Take the black shiny rectangle phone to the dark green holder with clamp.", "image": [ "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000000.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_003/sample_frames/head_camera/000128.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_015/sample_frames/head_camera/000075.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_015/sample_frames/left_camera/000075.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_015/sample_frames/right_camera/000075.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_015/sample_frames/head_camera/000060.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_015/sample_frames/left_camera/000060.png", "images/robotwin2_agilex_part1/place_phone_stand_aloha-agilex_randomized_500/episode_015/sample_frames/right_camera/000060.png" ], "conversations": [ { "from": "human", "value": "You are a rigorous and impartial vision evaluator specializing in physical manipulation tasks. Your goal is to determine the physical progress of a task by comparing two sets of observations.\n\n\n`Take the black shiny rectangle phone to the dark green holder with clamp.`\n\nREFERENCE EXAMPLES (For anchoring the full task scope):\n- REFERENCE START (Task just beginning): \n- REFERENCE END (Task fully completed): \n\n\nBEFORE Observation (Front, Left Wrist, Right Wrist):\n \n\nAFTER Observation (Front, Left Wrist, Right Wrist):\n \n\n### Goal\nCompare the BEFORE and AFTER image sets and judge whether AFTER is physically closer to the task objective than BEFORE, using the REFERENCE images as conceptual anchors.\n\n### Evaluation Criteria (Apply across all views)\n1) **Task Alignment**: Evidence must be directly tied to the objective of `Take the black shiny rectangle phone to the dark green holder with clamp.`.\n2) **Completeness & Accuracy**: Evaluate the quality of the execution, including correct pose, stable contact, placement accuracy, orientation, grasp quality, and the absence of collisions.\n3) **View-Specific Logic & Consistency**:\n - **Front View**: Use for global layout, object pose, approach path, and scene-level constraints.\n - **Wrist Views**: Use for **fine-grained details** such as gripper/hand engagement, contact area, slippage, micro-misalignment, object deformation, or unintended contact.\n - **Override Rule**: When views disagree, prioritize the view providing **decisive cues** (e.g., wrist views often override the front view for grasp validity or stable contact). \n - **Failure Rule**: If ANY single view shows a failure that invalidates success (e.g., a mis-grasp, a collision, or an unstable pose), this must be judged as regression.\n4) **Robustness**: Ignore irrelevant factors like lighting changes, color shifts, background clutter, or UI/watermarks that do not affect the physical task state.\n\n### Decision Logic\nYou must provide a binary judgment. No intermediate or zero scores are allowed:\n- **Output +1**: If the AFTER state shows clear progress, is more stable, or is physically closer to the objective than the BEFORE state.\n- **Output -1**: If the AFTER state shows regression, an error, a collision, or is physically further from the objective than the BEFORE state.\n\n### Output Format (STRICT)\nReturn ONLY the decision wrapped in tags:\n+1 or -1\n" }, { "from": "gpt", "value": "-1" } ], "hop_value": -20.0, "hop_absolute_value": 20.0, "hop_category": { "absolute_category": "small", "direction": "regression", "combined_category": "regression_small" }, "image_dataset": "robotwin2_agilex_part1" } ]